NemoDeepAgents CLI Commands Reference
The nemo-deepagents alias is the primary interface for managing Deep Agents sandboxes through NemoClaw.
It is installed automatically by the installer (curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_AGENT=langchain-deepagents-code bash).
Most commands in this reference use the same arguments and subcommands across agent variants.
Use nemo-deepagents when you want Deep Agents selected by default.
For guidance on choosing between the agent CLIs and the underlying openshell CLI, refer to CLI Selection Guide.
Agent Selection
Use nemo-deepagents for the Deep Agents variant.
It selects langchain-deepagents-code by default during onboarding and for other commands.
Use --agent langchain-deepagents-code, --agent dcode, or NEMOCLAW_AGENT=langchain-deepagents-code when you need the same selection through another entry point.
Deep Agents-specific sections below describe the dcode terminal runtime, managed /sandbox/.deepagents config, and commands that launch the interactive TUI or headless runner.
In-Sandbox Commands
Deep Agents does not use the OpenClaw chat slash command.
Use the host-side nemo-deepagents commands for lifecycle, status, policy, and inference operations.
Inside the sandbox, use dcode for the interactive TUI and dcode -n for explicit headless automation.
Add --json when automation needs the managed, versioned result envelope.
For the JSON schema, status and exit behavior, and 1 MiB output limit, refer to Run Deep Agents Code.
Hosted Installer Options
The hosted installer accepts options after bash -s --.
These options control installation and the onboarding run that follows it.
--local-model-runtime=vllm
Enable the fixed vLLM local model profile.
The flag accepts only vllm.
It makes the remaining onboarding non-interactive and disables Express profile selection.
The profile selects a fixed catalog model and serving command from the managed-inference catalog.
The hosted installer rejects NEMOCLAW_PROVIDER and NEMOCLAW_MODEL before onboarding.
The dedicated vLLM onboarder accepts NEMOCLAW_VLLM_MODEL only when the catalog resolves it to the matching fixed recipe.
It rejects a model that does not resolve to that recipe and all NEMOCLAW_VLLM_EXTRA_ARGS_JSON values before it installs vLLM.
Set NEMOCLAW_VLLM_PORT before installation to publish the fixed serving recipe on another host port.
The hosted installer’s equivalent environment-variable form requires both NEMOCLAW_ENABLE_LOCAL_MODEL_PROFILE=1 and NEMOCLAW_LOCAL_MODEL_RUNTIME.
Use the installer flag unless an automation boundary cannot pass installer arguments.
For prerequisites, effects, verification, and recovery, refer to Choose a Local Inference Server.
Hosted Installer Exit Statuses
The hosted installer reports how a run stopped through its exit status.
When you interrupt it at a prompt, it exits 130, the same status that nemo-deepagents onboard reports for that interrupt.
An interrupted onboarding run still prints [ERROR] Onboarding did not complete successfully. before it exits, so read the exit status rather than that line.
The installer preserves no other signal status, and a progress step stopped by SIGTERM also exits 130, so a script that stops the installer itself cannot read 130 as a deliberate interrupt.
DGX Station host preparation exits 10 when it requires a reboot and 11 when it requires a new login session, and prints the command that resumes the install.
Treat every other non-zero status as a failure.
Standalone Host Commands
The CLI handles host-side operations that run outside the selected agent runtime.
nemo-deepagents help, nemo-deepagents --help, nemo-deepagents -h
Show the top-level usage summary and command groups.
Running nemo-deepagents with no arguments shows the same help output.
nemo-deepagents --version, nemo-deepagents -v
Print the installed NemoClaw CLI version.
nemo-deepagents completion
Generate a tab-completion script for Bash, Zsh, or Fish from the commands and flags available in the installed CLI.
The script completes public global commands, the sandbox-first nemo-deepagents <name> ... grammar, flags, shell choices, and locally registered sandbox names.
If you omit the shell name, nemo-deepagents completion detects the target from $SHELL and defaults to Bash when it cannot identify Zsh or Fish.
The generated script is bound to the CLI name that created it, so install a separate script for each CLI alias you use.
It loads sandbox names from the local registry the first time completion runs and caches them for the rest of that shell session.
For Bash, source the generated script and add the same line to ~/.bashrc for future sessions.
For Zsh, source the generated script and add the same line to ~/.zshrc for future sessions.
For Fish, write the generated script to Fish’s completions directory.
Start a new shell session to refresh the cached sandbox names after creating or removing a sandbox.
nemo-deepagents resources
Display host hardware inventory and configured sandbox resource profiles.
Use --json for machine-readable CPU, memory, GPU, Kubernetes allocatable-capacity, and profile data.
If the gateway is not running, Kubernetes allocatable fields are omitted and host CPU/RAM totals are still shown.
nemo-deepagents host probe
Inspect host capabilities and gateway authority before onboarding without changing host, Docker, gateway, credential, policy, or sandbox state.
Use --json for the schema-versioned report.
The command exits with 0 for supported, 2 for incompatible, and 3 for inconclusive.
For capability IDs, evidence bounds, and compatibility guidance, refer to System Readiness.
nemo-deepagents agents list
List the installed agent runtimes that can be selected with nemo-deepagents onboard --agent <name>.
Use this global command when you need valid runtime names before creating or recreating a sandbox.
It lists runtime names with the descriptions from their installed manifests.
Expected output:
nemo-deepagents profiles list
List the serving profiles installed with NemoClaw and evaluate them against the current host. The command reports each profile’s stable ID, display name, inference backend, model, topology, selection mode, support state, estimated downloads, and incompatibility reason. It reads the serving catalog and host readiness state without downloading a model or changing host, gateway, inference, or sandbox resources.
Use --json for machine-readable output with the same profile fields.
Use the stable id value with nemo-deepagents onboard --profile <name>.
Display names are accepted when they identify exactly one profile, but stable IDs are suitable for scripts and automation.
nemo-deepagents onboard
Run the interactive setup wizard (recommended for new installs). The wizard creates an OpenShell gateway, registers inference providers, builds the sandbox image, and creates the sandbox. Use this command for new installs and for recreating a sandbox after changes to policy or configuration.
For Deep Agents, use the alias or pass the agent explicitly:
--agent accepts the canonical manifest names from nemo-deepagents agents list plus common aliases.
For example, nemohermes resolves to hermes, while dcode, deepagents, deepagents-code, and langchain resolve to langchain-deepagents-code.
--profile <name>
Select one serving profile from nemo-deepagents profiles list for interactive or non-interactive onboarding.
The flag is generic and does not add a model-specific command or flag.
NemoClaw maps a unique display name to its stable catalog ID and passes that ID to the managed inference path.
NemoClaw rejects an unknown, ambiguous, disabled, or incompatible profile before image or model downloads begin.
It also rejects --profile when you combine it with NEMOCLAW_PROVIDER, NEMOCLAW_MODEL, NEMOCLAW_VLLM_MODEL, NEMOCLAW_MANAGED_CLUSTER_PEERS, or NEMOCLAW_VLLM_EXTRA_ARGS_JSON overrides.
If NEMOCLAW_SERVING_PRESET is already set, it must select the same stable profile ID; a different ID conflicts with --profile.
Run nemo-deepagents profiles list to inspect an incompatibility reason before onboarding.
If you omit --profile, onboarding uses the same provider and model defaults as an installation without this feature.
The onboarding review screen identifies the resolved profile, recipe, declared model, served model alias, runtime image, support state, and download estimates before confirmation.
When onboarding reuses a running vLLM server, its /v1/models response must match the requested profile’s served alias or declared model root.
Otherwise, onboarding stops before it records a route that the profile does not declare.
After creation, human status shows the profile, recipe, and catalog digest; JSON status includes the complete secret-free servingProfileProvenance record for diagnostics and automation.
--host-mount
On Linux and Windows Subsystem for Linux 2 (WSL2), repeat --host-mount <absolute-host-directory:/sandbox/directory> to expose existing host directories read-only inside the sandbox.
The option requires a NemoClaw-managed Docker-driver gateway and does not provide a read-write mode.
Refer to Mount a Host Directory for Read-Only Access for validation rules, security considerations, persistence, and verification.
--events=jsonl
Emit a read-only stream of canonical onboarding FSM events as JSON Lines on stdout. Each line is one JSON object with the version 1 envelope:
In this mode, human progress remains available on stderr so stdout stays valid JSONL.
Payloads contain only the existing bounded, redacted machine-event context: credential environment variable names may appear, but credential values and secret-bearing URL components are redacted.
For a compatible-endpoint route that uses openai-completions, the context includes reasoningEffort as low, medium, high, or endpoint-default.
Other provider and API-family routes omit this field.
Treat new event type values and new payload fields as additive changes.
A breaking envelope or field-semantics change increments schemaVersion.
This surface observes the canonical onboarding session and does not accept input, cancel onboarding, or create another state machine.
It does not provide event history, reconnect, or replay; use the existing --resume behavior after an interrupted onboarding process.
Closing the output pipe or applying sustained backpressure disables observation without cancelling, rolling back, or otherwise changing onboarding.
Without --events=jsonl, terminal output and behavior are unchanged.
--resume and --fresh
NemoClaw records onboarding progress so interrupted runs can continue.
Use --resume to continue a resumable onboarding session with the provider, model, sandbox name, agent, observability choice, custom Dockerfile path, read-only host-mount declarations, and any explicitly selected serving-profile provenance recorded by the original run.
For a profile-backed session, resume requires the same catalog, preset, and recipe digests and exits before effects if the installed definition changed.
Omit --profile to reuse that recorded selection, or pass the same profile explicitly; use --fresh to adopt a changed catalog definition.
Sessions without a serving-profile provenance record can resume when their checkpoint uses schema 4, but they cannot acquire a new --profile selection during resume.
Checkpoint schema 4 records whether onboarding uses the default profile or the portable experimental profile.
For the portable profile, it also records the current user’s canonical home reported by the operating system, that home’s exact .config directory, the runtime root, rootless Podman endpoint path, and runtime ownership.
It does not record ambient Docker or Podman runtime selector values.
The runtime authority record contains no credentials.
A plain --resume restores the recorded profile.
You can also run nemo-deepagents onboard --experimental-profile portable --resume when the recorded profile is portable.
NemoClaw rejects an explicit profile that conflicts with the checkpoint before it changes portable configuration, activates the user-scoped Podman socket, or changes gateway and sandbox resources.
Portable resume derives DOCKER_HOST, CONTAINERS_CONF, and NETAVARK_FW again while it holds the onboarding lock.
It ignores ambient Docker and Podman runtime selectors during that derivation.
NemoClaw scopes the derived values to onboarding and restores the process environment after success or failure.
It verifies the current user, canonical roots, socket path and ownership, Podman identity and version, and required configuration before a resumed onboarding step changes resources.
Resume stops before writes or activation if an existing socket or configuration path is a symlink, has the wrong owner, or has an unsafe type or mode.
NemoClaw can create missing descendants beneath a validated current-user root and reconcile content drift in its own portable configuration files.
A missing user-scoped socket after a host reboot can be activated and verified at the recorded path.
A new socket inode or a supported Podman upgrade does not invalidate the checkpoint.
Portable onboarding always uses the .config directory beneath the canonical home reported by the operating system.
HOME and XDG_CONFIG_HOME never select or override this authority.
NemoClaw ignores ambient XDG_CONFIG_HOME during onboarding and restores its exact prior presence and value afterward.
Resume rejects a checkpoint that records another configuration root.
It also rejects stored authority or filesystem ownership drift without falling back to Docker.
Checkpoint Resume Compatibility
An active onboarding session with checkpoint schema 1, 2, or 3 cannot resume because those schemas did not record the default or portable profile authority.
NemoClaw preserves the older session and exits before portable configuration, socket activation, or resource changes.
Run nemo-deepagents onboard --fresh to discard the active session and start fresh onboarding.
If you intend to use the portable experimental profile, run nemo-deepagents onboard --experimental-profile portable --fresh.
This compatibility restriction does not prevent NemoClaw from reading a completed older session during status inspection.
Before the configuration review, NemoClaw records the sandbox name and the selected provider and model as an incomplete choice.
If onboarding stops at the review prompt, an interactive --resume run shows the prompt again.
A non-interactive --resume run reuses the recorded choice and continues to inference setup.
After you choose Apply configuration, NemoClaw records the choice before inference setup starts.
If inference setup fails, --resume reuses the accepted provider, model, and sandbox name.
If you choose Exit onboarding, onboarding exits with a nonzero status and clears those recorded choices.
Run nemo-deepagents onboard to make new choices after exit.
During a resume without terminal input, --yes or NEMOCLAW_YES=1 also selects non-interactive resume behavior.
For a new or fresh session, --yes and NEMOCLAW_YES=1 accept supported confirmations but do not replace --non-interactive.
If onboarding returns without reaching the final complete state, the command exits with status 1.
When that result is resumable, NemoClaw keeps the session in_progress at its last checkpoint instead of marking it failed, so correct the reported condition and run nemo-deepagents onboard --resume.
Completed onboarding sessions are not resumable.
Use --resume only for resumable interrupted or failed sessions, not to change provider, model, agent, or sandbox recreation settings after onboarding has completed.
During resume, NemoClaw reruns preflight, gateway, provider, and sandbox repair checks even when the saved session has already reached a later nonterminal onboarding phase.
If the recorded session conflicts with flags you pass on the recovery run, NemoClaw exits and tells you to either rerun with the original settings or start over.
An active same-name replacement is separate from ordinary onboarding-step resume.
If onboarding printed Journaled replacement before it stopped, rerun the original onboarding command with the same target settings.
The replacement can continue without an explicit --resume flag.
Refer to Continue an Interrupted Replacement for the identity checks and failure conditions.
Use --fresh to discard the saved onboarding session and start the wizard from the beginning.
This clears stale or failed session state before NemoClaw creates a new session record.
It also bypasses locally recorded sandbox base-image resolution metadata and reruns normal candidate resolution.
--fresh takes precedence over a base-image hint carried from a rebuild, so NemoClaw does not use that recorded hint.
The installer also accepts --fresh and forwards it to nemo-deepagents onboard, which skips automatic resume detection.
--resume and --fresh are mutually exclusive.
For an existing completed sandbox, use --fresh --name <sandbox-name> --recreate-sandbox when you intentionally want onboarding to replace that sandbox with a new provider, model, agent, or build-time setting.
Use nemo-deepagents <sandbox-name> rebuild when you want NemoClaw to recreate the sandbox from its recorded registry metadata without changing those selections.
--tool-disclosure <progressive|direct>
Choose how the selected agent presents its session-authorized tools to the model.
progressive is the default: OpenClaw and Hermes use their native Tool Search implementations, while Deep Agents Code initially shows its core tools plus search_tools after at least one MCP tool loads successfully.
direct restores the previous behavior and presents all registered tools directly.
This setting changes model context only; it does not bypass OpenShell policy, credentials, approvals, hooks, or sandbox controls.
The flag takes precedence over NEMOCLAW_TOOL_DISCLOSURE.
A new sandbox defaults to progressive when neither is set.
NemoClaw records the selected value with the onboarding session and sandbox so rebuilds preserve it and ambient shell variables cannot silently change an internal rebuild.
Model-specific compatibility safeguards may downgrade a selected progressive mode to direct exposure for that model without changing the recorded preference.
To change an existing sandbox, recreate it explicitly:
Without an explicit flag or environment value, recreation preserves the recorded setting and only falls back to progressive for legacy state.
Resuming an interrupted session with a different explicit setting fails with a conflict instead of changing behavior mid-session.
--observability and --no-observability
Enable backend-neutral trace export for a LangChain Deep Agents Code sandbox.
During initial onboarding, pass --observability with the Deep Agents alias.
When you use the generic nemo-deepagents entry point, combine it with --agent langchain-deepagents-code.
NemoClaw rejects the positive flag for OpenClaw and Hermes sandboxes.
Use --no-observability when you need to clear a recorded Deep Agents Code choice before switching the resumed session to another agent.
The flag is off by default.
When enabled, NemoClaw records the choice with the onboarding session and sandbox, adds the observability-otlp-local policy preset on supported policy tiers, and preserves the choice across resume and rebuild operations.
An explicit --observability or --no-observability choice updates a resumed onboarding session.
The Restricted tier suppresses automatic application of the preset.
An operator can add it manually after reviewing the additional egress, but the next Restricted onboarding or rebuild reconciliation removes it.
The explicit opt-in can export bounded prompts, responses, tool arguments, tool results, and operational metadata.
Treat trace payloads as sensitive application data.
The managed capture applies size, depth, item-count, recognized-key, and exception-text safeguards, but it does not detect secrets embedded in ordinary content values.
Deep Agents Code sends OTLP/HTTP protobuf traces to the fixed local endpoint http://host.openshell.internal:4318/v1/traces.
The OTLP library adds standard transport headers, but the sandbox cannot configure operator-supplied custom or authentication headers, a remote endpoint, backend credentials, or a backend.
Changing this setting on an existing sandbox requires a new sandbox process so the startup environment matches the recorded choice. Use the transactional rebuild flags so NemoClaw backs up declared agent state, preserves managed MCP providers and adapter state, recreates the sandbox, and restores the backup.
Removing the observability-otlp-local policy stops delivery immediately but does not clear the recorded opt-in.
A later rebuild restores the preset on Balanced and Open tiers, while Restricted continues to suppress it.
For policy recovery and the host-side LangSmith exporter example, refer to Set Up Deep Agents Trace Export.
Review Understand Deep Agents Trace Export for the privacy boundary, Verify Deep Agents Trace Export for delivery checks, and Manage Deep Agents Trace Export for lifecycle operations.
When Docker exposes the required identity metadata, NemoClaw records the base-image resolution on managed sandbox images.
During a warm recreate or rebuild, it validates the local image identity and platform, plus the exact repository digest for a published image and any active OpenShell ABI requirement, before reusing it.
A valid match avoids candidate discovery and a network pull.
Set NEMOCLAW_SANDBOX_BASE_IMAGE_REFRESH=1 to bypass the recorded hint without changing onboarding session handling:
Base-image selection follows this precedence:
--freshorNEMOCLAW_SANDBOX_BASE_IMAGE_REFRESH=1bypasses recorded metadata and reruns normal candidate resolution. These controls are equivalent for base-image selection.- Without a bypass, NemoClaw validates and reuses the recorded hint when possible.
- When the hint is absent or no longer valid, NemoClaw performs normal resolution.
After a cache miss, source checkouts require a fresh local build before candidate selection when base-image inputs have dirty or staged changes, or Git cannot inspect the worktree safely.
For a clean release checkout or versioned install, NemoClaw first accepts the exact release-version image.
If that tag exists locally but fails compatibility validation, NemoClaw refreshes the same tag from the registry once and validates it again.
If the release-version image is missing or still incompatible, NemoClaw builds a compatible local base instead of falling back to mutable :latest.
For clean unversioned development checkouts, NemoClaw first tries the image tagged with the exact source commit.
If that image is unavailable and committed base-image inputs differ from main, NemoClaw requires a compatible local build.
When committed base-image inputs match main, NemoClaw tries the image tagged with the newest reachable release version from origin and only uses :latest when no version tag is discoverable.
When a stable tag and a prerelease tag share the same version, NemoClaw prefers the stable tag.
If origin tag lookup is unavailable, NemoClaw uses the newest reachable local release tag as a fallback.
If that nearest release-version image is missing or incompatible, NemoClaw builds a compatible local base instead of falling back to mutable :latest.
The required-build path does not reuse an older local tag.
If local builds are disabled or the build fails, resolution stops instead of selecting a stale image.
When the OpenShell sandbox ABI is required, NemoClaw also rejects a built image that does not report a compatible glibc version.
Explicit base-image overrides are exact: NemoClaw validates the requested ref and fails closed when it cannot be pulled or does not satisfy required ABI, agent runtime, or dependency checks.
Otherwise, normal resolution checks compatible images in Docker’s local image store before attempting to pull a missing published candidate.
For warm-hint reuse and unversioned development resolution, NemoClaw can reuse another validated local fallback when published candidates are unavailable or incompatible.
When the OpenShell sandbox ABI is required, that local fallback must be ABI-compatible.
An offline warm recreate or rebuild can therefore continue when the recorded image or another compatible candidate is available locally.
When source inputs require a fresh local build, NemoClaw fails the operation if that build cannot be produced and validated instead of substituting an older local tag.
When the OpenShell sandbox ABI is required, resolution also fails if no ABI-compatible image can be resolved instead of falling back to an unvalidated cached :latest image.
Bypassing the recorded hint does not clear Docker’s local image store or require a network pull.
Only --fresh also discards the saved onboarding session; the refresh environment variable affects base-image selection only.
For NemoClaw-managed environments, use nemo-deepagents onboard when you need to create or recreate the OpenShell gateway or sandbox.
Avoid openshell self-update, npm update -g openshell, or openshell sandbox create directly unless you intend to manage OpenShell separately and then rerun nemo-deepagents onboard.
Use --fresh to ignore any saved onboarding session and restart the wizard from scratch. This is useful after an interrupted nemo-deepagents onboard run when you want to discard saved state instead of continuing it with --resume.
The installer detects existing sandbox sessions before onboarding and prints a warning if any are found.
To make the installer abort instead of continuing, set NEMOCLAW_SINGLE_SESSION=1:
When existing sandboxes were created with OpenShell earlier than 0.0.37, the installer prompts before running the automatic gateway upgrade path.
For scripted installs, set NEMOCLAW_ACCEPT_EXPERIMENTAL_OPENSHELL_UPGRADE=1 to allow the automatic path to prepare the current CLI without replacing OpenShell, back up every registered sandbox with the current state manifest, retire an installed gateway whose OpenShell version is outside the current release’s supported range, install the supported OpenShell release, and recover the existing sandboxes.
The installer reads that supported range from the prepared current source and stops without retiring the gateway if the installed version is unknown or the range is missing or invalid.
When the installed OpenShell version is already supported, the installer keeps the running gateway through the host update.
On Linux, if installed OpenShell lifecycle commands cannot retire the gateway, the installer checks a verified NemoClaw-managed gateway PID file for any configured gateway port.
For the default gateway on port 8080, the installer first checks a verified active nemoclaw-openshell-gateway.service, then checks the PID file.
After either fallback confirms the gateway process is stopped, the installer tries to remove the selected OpenShell registration and warns if onboarding must replace a stale registration.
If neither fallback can verify and stop the process, the installer stops after backup with every sandbox backup preserved.
If any registered sandbox cannot be backed up, the installer aborts before it changes the gateway.
After the automatic path retires an out-of-range gateway, it forces installation of the OpenShell version pinned by the prepared source before recovery.
This mandatory installation applies to source and managed install modes and cannot remain deferred after gateway retirement.
If the forced installation fails, the installer does not stage a gateway service or start recovery, preserves the backups, and tells you to rerun with NEMOCLAW_OPENSHELL_UPGRADE_PREPARED=1.
When the registry contains a pre-fingerprint OpenClaw or Hermes entry with no recorded custom-image evidence, an interactive install asks you to confirm that the listed sandbox used a NemoClaw-managed image.
For a non-interactive install, set NEMOCLAW_CONFIRM_LEGACY_MANAGED_RECREATE to the exact JSON array of names printed by the installer, such as ["my-assistant","preserve-hermes"], only after verifying every named sandbox used a managed image.
The confirmation permits those legacy entries to recover onto the current managed image, but it does not override recorded custom-image evidence.
After successful recovery, the installer skips generic onboarding.
For any registered-sandbox upgrade that you already prepared manually, set NEMOCLAW_OPENSHELL_UPGRADE_PREPARED=1 only after backing up every registered sandbox and retiring the old gateway.
This environment variable asserts that those steps are complete, so the installer skips the repeated backup and gateway-retirement phase before it checks whether OpenShell is installed or whether its version is in range.
For a non-default gateway, preserve the selected port on the bash side of the install pipeline.
It reuses the latest backups, forces the pinned OpenShell installation, and starts recovery only after that installation succeeds.
If the installation fails, rerun the same install-pipeline command to preserve NEMOCLAW_GATEWAY_PORT and NEMOCLAW_OPENSHELL_UPGRADE_PREPARED.
Legacy Upgrade Recovery Scope
Prepared backup recovery for a legacy sandbox restores only the managed state directory recorded in its validated manifest, such as /sandbox/.openclaw or /sandbox/.hermes.
Files outside that recorded path, including /sandbox/user-data, are not preserved when the installer recreates the sandbox.
Back up those paths outside the sandbox before you continue.
The wizard prompts for a provider first, then collects the provider credential if needed.
Supported non-experimental choices include NVIDIA Endpoints, OpenRouter, OpenAI, Anthropic, Google Gemini, and compatible OpenAI or Anthropic endpoints.
Credentials are registered with the OpenShell gateway and never persisted to host disk.
Refer to Credential Storage for details on inspection, rotation, and migration from earlier releases.
The legacy nemo-deepagents setup command is deprecated; use nemo-deepagents onboard instead.
On a qualified DGX Spark, the provider menu lists compatible experimental managed llama.cpp profiles in descending YAML priority order.
During interactive onboarding without an explicit provider request, the menu ignores NEMOCLAW_LLAMACPP_RECIPE and marks the unique highest-priority compatible profile as (recommended).
The recommended profile appears as Managed llama.cpp: Meta Muse Glimmer 30B on one DGX Spark (recommended).
The NVIDIA Nemotron profile appears next without the recommendation marker.
The selected menu entry determines the exact recipe even when NEMOCLAW_LLAMACPP_RECIPE names another recipe.
Select the same path non-interactively with the repository-owned recipe:
Use llama-cpp.nemotron-3-nano-30b-a3b.spark-single.v1 to select the lower-priority NVIDIA Nemotron recipe explicitly.
Do not set NEMOCLAW_MODEL for the managed llama.cpp path.
For prerequisites, external traffic, verification, and recovery, refer to Install Managed llama.cpp on DGX Spark.
After provider selection, the wizard reviews the provider, model, credential state, and sandbox name before registering inference. The interactive review offers these actions:
- Apply configuration continues to provider registration.
- Edit inference provider or model returns to provider and model selection.
- Edit sandbox name prompts for the sandbox name again.
- Exit onboarding stops onboarding before provider registration.
When you edit inference, NemoClaw clears the credential staged for the discarded selection.
NemoClaw preserves the sandbox name.
When you edit the sandbox name, NemoClaw preserves the inference selection.
The sandbox prompt shows the prior name as its default.
After you apply the configuration, routine editing ends.
If inference setup fails and offers a back recovery action, you can return to provider and model selection and then review the updated configuration again.
It then prompts for optional web search, builds and starts the sandbox, and asks for a policy tier that controls the default set of network policy presets applied to the sandbox.
Four tiers are available:
After selecting a tier, the wizard shows a combined preset and access-mode screen where you can include or exclude individual presets and toggle each between read and read-write access.
When Personal is selected or carried forward, personal-open-internet is mandatory for every agent and every onboarding entry point, including Portable.
The picker and policy modes control only additional presets; they cannot deselect, skip, or replace Personal’s required web authority.
For details on tiers and the presets each includes, refer to Network Policies.
When you finish the policy step, NemoClaw records the finalized built-in preset selection for that sandbox.
When onboarding creates or recreates a sandbox with presets, NemoClaw prints the exact finalized create-time policy scope before registering providers or creating the sandbox.
Later re-onboard runs seed from that finalized selection, so presets you intentionally removed stay removed unless you select them again or override the policy mode.
In non-interactive mode, set the tier with NEMOCLAW_POLICY_TIER (default: balanced):
Unset, blank, or whitespace-only NEMOCLAW_POLICY_TIER values use the balanced default.
In non-interactive mode, any non-blank value must be one of restricted, balanced, open, or personal; otherwise onboarding exits before preflight, gateway, or inference side effects with an error listing the valid options.
Interactive onboarding ignores an invalid environment value and shows the normal tier prompt.
NEMOCLAW_POLICY_MODE controls how non-interactive onboarding reconciles the tier-derived suggestions against the sandbox’s currently-applied presets.
The default is suggested, which is additive.
Onboarding applies tier defaults and preserves any presets you previously added with nemo-deepagents <name> policy add across re-onboards.
Use custom with NEMOCLAW_POLICY_PRESETS when you want the explicit list to be authoritative for optional presets.
Onboarding removes any optional preset that is not in the list.
skip does not add optional tier defaults and retains eligible optional presets already applied.
For Personal, all modes still apply or retain the mandatory personal-open-internet preset.
NemoClaw filters tier suggestions and resume selections by active agent support and the selected web search provider.
During automatic suggestion and resume reconciliation, it removes stale web-search selections when they conflict with the active agent or selected provider.
The Personal tier instead uses personal-open-internet for web transport and does not select Brave Search or Tavily Search merely to enable ordinary web fetches.
This makes keyless fetches available to any sandbox binary, but it does not add a provider-free web_search implementation.
An explicit custom preset list or interactive manual selection remains operator-controlled for additional presets.
Deep Agents onboarding supports the maintained Tavily Search path.
NemoClaw registers the Tavily credential with the OpenShell gateway, applies the tavily policy preset when you opt in, and rebuilds the sandbox so the provider attaches to the managed Python runtime.
Do not place TAVILY_API_KEY in /sandbox/.deepagents/.env, .state/auth.json, or other Deep Agents Code state.
For non-interactive onboarding, export the Tavily key only in the host shell that runs onboarding:
For non-interactive onboarding, you must explicitly accept the third-party software notice:
or:
For scripted installer runs, pass explicit acceptance to the bash side of the installer pipe:
If the installer cannot prompt for the notice in a terminal and no explicit acceptance is set, it exits before installing Node.js or the NemoClaw CLI.
The wizard prompts for a sandbox name.
Names must contain 1 to 19 characters.
They must be lowercase, start with a letter, contain only letters, numbers, and single internal hyphens, and end with a letter or number.
Consecutive hyphens (--) are not allowed.
The CLI rejects names that do not match these rules.
It also prints a Try: <suggested-slug> recovery line whenever it can derive a valid lowercase, hyphen-separated form from the input, so passing --name MyAssistant reports Try: myassistant.
Names that match global CLI commands (status, list, debug, etc.) are rejected to avoid routing conflicts.
Use --agent <name> to target a specific installed agent profile during onboarding.
The nemo-deepagents onboard --help output lists installed runtime names inline, and nemo-deepagents agents list shows the same runtimes with manifest descriptions.
If you cancel a brand-new onboarding run at the policy preset step, NemoClaw rolls back the sandbox, registry entry, and onboarding session instead of leaving a default sandbox with unfinished policy state. Existing live sandboxes are not deleted by this cancel rollback path.
If you run onboarding again with the same sandbox name and choose a different inference provider or model, NemoClaw detects the drift and recreates the sandbox so the running agent config matches your selection.
In interactive mode, the wizard asks for confirmation before delete and recreate.
In non-interactive mode, NemoClaw recreates automatically when the stored selection is readable and differs.
For managed Deep Agents Code sandboxes, NemoClaw also recreates when the live dcode identity selection is unreadable; other agent paths continue to reuse by default when their stored selection cannot be read.
Set NEMOCLAW_RECREATE_SANDBOX=1 to force recreation even when no drift is detected.
Before deleting an existing sandbox during recreation, NemoClaw backs up the workspace state declared by the selected agent profile and restores it into the new sandbox once it is live.
This applies whether the existing sandbox is ready or marked not-ready, so cross-version upgrades that pass NEMOCLAW_RECREATE_SANDBOX=1 no longer drop user files from the selected agent workspace.
The behaviour matches nemo-deepagents <name> rebuild --force.
NemoClaw aborts the recreate when the backup cannot complete in full, including when individual state directories or files fail mid-backup, so failed entries are not silently dropped on delete.
Set NEMOCLAW_RECREATE_WITHOUT_BACKUP=1 to skip the pre-recreate backup.
The destination sandbox starts with a fresh workspace.
Before deletion, onboarding prints a Journaled replacement diagnostic with the replacement identifier, recorded OpenShell gateway, and current phase.
If the process stops after this point, a later same-target onboarding run continues the active replacement without requiring --resume.
It accepts a ready same-name replacement only when the live identity and sandbox registry generation match the journal.
It fails closed if the gateway, source, target, durable source registry fields, or replacement settings changed.
Before creating the gateway, the wizard runs preflight checks.
It verifies that Docker is reachable and prints host remediation guidance when prerequisites are missing.
Standard onboarding rejects unsupported runtimes such as Podman.
The explicit portable experimental profile has one installer-preflight admission exception for the Podman unsupported-runtime finding.
It does not waive any other readiness blocker or make Podman generally supported.
The preflight also enforces the OpenShell version range declared in the blueprint (min_openshell_version and max_openshell_version).
If the installed OpenShell version falls outside this range, onboarding exits with an actionable error and a link to compatible releases.
For fresh OpenShell installs, NemoClaw queries published OpenShell releases and asks the installer to use a release that fits the blueprint range.
If release metadata is unavailable, the installer uses its bundled fallback pin and the post-install version gate still enforces the range.
When NemoClaw finds an existing gateway to reuse, it probes the host gateway HTTP endpoint before declaring the gateway reusable.
If the container is running but the upstream is still warming up (for example, immediately after a Docker daemon restart), NemoClaw rebuilds the gateway instead of trusting stale metadata.
On the Docker-driver gateway path, preflight stays read-only when it detects a stale gateway (for example, a Docker-driver runtime env hash drift).
It prints a ⚠ Gateway will be recreated when sandbox creation starts notice and defers the actual teardown to step [2/8] Starting OpenShell gateway.
This means pressing Ctrl+C between preflight and step [2/8] leaves the running gateway and existing sandbox containers untouched, so nemo-deepagents onboard is safe to run just to check preflight output.
An interrupted run prints the resume command and exits with status 130 for Ctrl+C or 143 for SIGTERM.
For Linux Docker-driver gateways, onboarding also checks that a helper container on the OpenShell Docker network can reach host.openshell.internal:<gateway-port>.
If a host firewall blocks that sandbox path, onboarding exits with a sudo ufw allow from <subnet> to <gateway-ip> port <gateway-port> proto tcp command before it reports the gateway healthy.
Set NEMOCLAW_AUTO_FIX_FIREWALL=1 to opt in to automatic UFW remediation for this specific failure: NemoClaw uses sudo -n only, validates the Docker bridge subnet/gateway/port, applies the narrow UFW rule only after a proven TCP reachability failure, and re-probes before continuing.
If passwordless sudo, UFW, or active UFW is unavailable, NemoClaw falls back to the manual guidance path without prompting for a password.
For the portable experimental profile, the helper maps host.openshell.internal to the OpenShell Podman host gateway instead of the inspected network gateway.
This path does not use Docker bridge UFW remediation.
After all portable TCP probe attempts fail, onboarding prints commands for the user-scoped Podman service and socket.
Onboarding prints the same commands when the portable probe cannot reach the user-scoped Podman service.
The printed rerun command keeps the portable experimental profile selected.
Portable commands reconstruct the current user’s rootless Podman socket authority from NemoClaw state before they use the Docker-compatible API.
They do not select an endpoint from ambient Docker or Podman runtime variables or named connections.
When podman.service reports inactive and the recorded socket exists, NemoClaw first makes one 10-second API request through the guarded recorded authority.
A valid server version classifies the endpoint as warm and avoids starting another socket service.
A missing socket or a response without a valid server version enters bounded cold activation.
Any socket authority change during this precheck fails at the socket authority stage.
When the user-scoped socket-backed service needs activation, NemoClaw activates it and waits through a bounded startup period for a real Podman API response.
During cold activation, the first API probe can cause systemd to replace the socket inode.
NemoClaw requalifies one such replacement and repeats the probe only when the socket path, device, mode, owner, and complete directory authority remain unchanged.
Any other authority change or a second inode replacement fails the readiness check.
After cold activation succeeds, later API health checks use the fixed 10-second steady-state deadline.
Onboarding and portable sandbox lifecycle commands use this same readiness contract.
Failures identify socket authority, service activation, startup API health, or steady-state API health without reporting credentials.
NemoClaw does not fall back to Docker or report an absent or unreachable endpoint as healthy.
A successful cold path uses the cold timing label and reports activation, API, and total time in milliseconds.
A successful warm path uses the warm timing label and reports steady-state API and total time in milliseconds.
To tune the existing-gateway HTTP health poll, use NEMOCLAW_REUSE_HEALTH_POLL_COUNT (default 6) and NEMOCLAW_REUSE_HEALTH_POLL_INTERVAL (default 5 seconds).
The poll count is clamped to a minimum of 1 so the health probe always runs at least once, and the interval is clamped to a minimum of 0 (no sleep between attempts).
--from <Dockerfile>
Build the sandbox image from a custom Dockerfile instead of the stock NemoClaw image.
The supplied Dockerfile defines the complete sandbox image, and NemoClaw does not layer it on top of the stock managed runtime.
The entire parent directory of the specified file is used as the Docker build context, so any files your Dockerfile references (scripts, config, etc.) must live alongside it.
When the supplied path is the selected agent’s own managed Dockerfile (for example, agents/hermes/Dockerfile in the NemoClaw checkout the CLI runs from), NemoClaw applies one exception and stages the repository root as the build context, exactly as the managed build does, because that Dockerfile copies repository-root paths.
This lets you edit the managed Dockerfile in place (for example to add Python packages) and rebuild from it with --from.
For this managed exception, onboarding applies the .dockerignore from the repository root.
For every other --from path, onboarding applies a .dockerignore from the Dockerfile’s parent directory while calculating the context size and staging files for Docker.
NemoClaw also applies additional secret-safety exclusions that override .dockerignore negation rules: credential-style files and directories such as .env*, .ssh/, .aws/, .netrc, .npmrc, secrets/, *.pem, and *.key are still skipped even if .dockerignore tries to include them.
Without a .dockerignore, onboarding still skips common large or local-only directories (node_modules, .git, .venv, and __pycache__) while staging this context.
Other build outputs such as dist/, target/, or build/ are included unless your .dockerignore excludes them.
If the staged context is larger than 100 MB, onboarding prints a warning before the Docker build starts.
Move the Dockerfile into a smaller dedicated directory or add .dockerignore entries for generated artifacts to shrink the context.
If the directory contains unreadable files (for example, Windows system files visible in WSL), onboarding exits with an error suggesting you move the Dockerfile to a dedicated directory.
NemoClaw builds user-supplied --from contexts with the OpenShell gateway builder.
The host-side local BuildKit prebuild is limited to build contexts generated entirely by NemoClaw.
On a local Docker-driver gateway, a Local BuildKit build skipped notice is expected and onboarding continues with the custom image.
The Dockerfile path must exist. Missing paths fail during command parsing before preflight, gateway setup, inference setup, or sandbox creation starts.
The file can have any name; if it is not already named Dockerfile, onboard copies it to Dockerfile inside the staged build context automatically.
To create an isolated build context, create a dedicated directory that contains only the Dockerfile and the files it needs:
For faster custom builds, plan for Docker cache behavior:
- Treat the first build on a fresh host as a cold build. Cold builds download the base image and package indexes, so they take longer than later warm rebuilds even when NemoClaw is healthy.
- A warm rebuild reuses cached layers when the base image and earlier layers are unchanged, so it is much faster than the first build.
- Order Dockerfile instructions from least-changing to most-changing: base image, system packages, dependency manifests, dependency install, then application source. This lets warm rebuilds reuse cached dependency layers instead of reinstalling on every source change.
- Pin the base image to an explicit tag or digest so warm rebuilds resolve the same cached base instead of pulling a new one.
To diagnose where a slow build spends time, set NEMOCLAW_TRACE=1 and read the phase timings in Onboard Profiling Traces.
NemoClaw does not guarantee exact build timings.
All NemoClaw build arguments (NEMOCLAW_MODEL, NEMOCLAW_INFERENCE_PROVIDER_ID, NEMOCLAW_INFERENCE_BASE_URL, etc.) are injected as ARG overrides at build time, so declare them in your Dockerfile if you need to reference them.
NEMOCLAW_INFERENCE_PROVIDER_ID is a non-secret inference route identifier (for example inference for proxied providers, or a provider family such as openai), never a credential; provider credentials stay in OpenShell provider storage.
It replaces the former NEMOCLAW_PROVIDER_KEY image argument, whose secret-shaped name triggered a BuildKit SecretsUsedInArgOrEnv warning.
The host-side NEMOCLAW_PROVIDER_KEY credential alias is unchanged; this migration only renames the managed image route selector.
Custom Dockerfiles that declare either ARG NEMOCLAW_INFERENCE_PROVIDER_ID or the legacy ARG NEMOCLAW_PROVIDER_KEY continue working in v0.0.91.
NemoClaw updates whichever supported declaration is present, and runtime consumers read the legacy name as a fallback.
Rename the legacy ARG/ENV declaration to NEMOCLAW_INFERENCE_PROVIDER_ID; the legacy fallback is retained for compatibility in this release and may be removed in a future release.
Custom Dockerfiles must declare ARG NEMOCLAW_TOOL_DISCLOSURE=progressive exactly once in the final build stage and promote it into that stage’s runtime environment.
The usual runtime contract is:
Onboarding and rebuild preflight reject a missing, duplicate, or unconsumed declaration before replacing an existing sandbox.
In non-interactive mode, the path can also be supplied via the NEMOCLAW_FROM_DOCKERFILE environment variable.
You must also supply a sandbox name via --name <sandbox> or NEMOCLAW_SANDBOX_NAME so a --from build cannot silently clobber the default my-assistant sandbox.
If a --resume is attempted with a different --from path than the original session, onboarding exits with a conflict error rather than silently building from the wrong image.
--name <sandbox>
Set the sandbox name without going through the interactive prompt.
The same name format and reserved-name rules that the wizard enforces apply here too.
Names must contain 1 to 19 characters.
They must be lowercase, start with a letter, contain only letters, numbers, and single internal hyphens, and end with a letter or number.
Consecutive hyphens (--) are not allowed.
Names that match a NemoClaw CLI command (status, list, debug, etc.) are rejected up front.
The flag wins over NEMOCLAW_SANDBOX_NAME.
When prompting is possible, NEMOCLAW_SANDBOX_NAME fills the interactive default so you can press Enter to accept it.
When prompting is impossible (no TTY or --non-interactive), the env var is also honoured so existing CI scripts keep working.
Combining --from <Dockerfile> with non-interactive onboarding requires one of --name or NEMOCLAW_SANDBOX_NAME; otherwise onboarding exits rather than silently defaulting to my-assistant and clobbering the default sandbox.
nemo-deepagents onboard --from
Use a custom Dockerfile for the sandbox image.
This variant of nemo-deepagents onboard accepts a --from <Dockerfile> argument to build the sandbox from a user-supplied Dockerfile instead of the default NemoClaw image.
The user-supplied context uses the OpenShell gateway builder instead of NemoClaw’s host-side local BuildKit prebuild.
GPU Passthrough
When nemo-deepagents onboard detects an NVIDIA GPU on the host, it enables OpenShell GPU passthrough at both the gateway and sandbox level by default.
The nvidia-smi probes require a successful result and reject placeholder JMJWOA-Generic-* GPU names unless NemoClaw can prove a supported NVIDIA platform or GPU execution.
NemoClaw treats a recognized NVIDIA product model from /sys/class/dmi/id/product_name or /sys/firmware/devicetree/base/model, or a known Tegra device node, as authoritative platform identity.
On eligible native or Docker Desktop-backed WSL ARM64 Linux hosts without that firmware evidence, one bounded Docker CUDA workload can prove GPU execution.
On those hosts, a single plausible, non-placeholder NVIDIA GPU name also requires that proof when the NVIDIA kernel-driver interface (/proc/driver/nvidia) is absent.
For Windows-on-Arm, this proof is a technical detection check and does not change the Unsupported product status or establish platform qualification.
Refer to Platform Support and Launch Claims for the current support boundary.
For the proof command, timeout control, and failure recovery, refer to GPU Setup Fails with a Placeholder GPU Name.
The names-only unified-memory fallback does not run this workload and rejects denylisted names.
Other non-firmware-vouched hosts also reject denylisted names.
Jetson/Tegra hosts that ship without nvidia-smi continue to be detected via the devicetree firmware fallback (/sys/firmware/devicetree/base/model) or the Tegra device-node fallback (/dev/nvhost-gpu, /dev/nvhost-ctrl-gpu, /dev/nvhost-ctrl, or /dev/nvmap); both bypass the trust-tier gate above.
Use --no-gpu to opt out when you want host-side inference providers only and do not need direct GPU access inside the sandbox.
Use --gpu to require GPU passthrough and fail fast if an NVIDIA GPU is not detected.
Use --sandbox-gpu or --no-sandbox-gpu to control only direct NVIDIA GPU access inside the sandbox.
Use --sandbox-gpu --sandbox-gpu-device <device> to pass a specific OpenShell GPU device selector to openshell sandbox create; device selectors require explicit sandbox GPU enablement.
On ordinary native Linux Docker-driver hosts, NemoClaw uses native OpenShell GPU injection by default and never broadens confinement automatically.
Portable onboarding requires native OpenShell GPU injection for every agent.
It does not use NEMOCLAW_DOCKER_GPU_PATCH compatibility routing, so do not set fallback, 1, or another legacy nonzero value for that profile.
Set NEMOCLAW_DOCKER_GPU_PATCH=fallback to explicitly authorize one native attempt followed by one compatibility retry.
NemoClaw permits the retry only after it confirms either a trusted host-side GPU routing failure or an explicit driver proof plus exact-container host configuration showing that no GPU was attached.
It then saves redacted diagnostics and removes the incomplete sandbox before retrying.
Sandbox-reported CUDA output alone never authorizes the broader compatibility envelope, even when the operator enabled fallback.
That case fails closed and points to the explicit NEMOCLAW_DOCKER_GPU_PATCH=1 compatibility-only control.
NemoClaw retries only after it verifies that no OpenShell-managed Docker container labeled for that sandbox remains; if cleanup cannot be proven safe, onboarding stops and prints cleanup guidance instead.
On Docker Desktop WSL and Jetson/Tegra, automatic GPU onboarding uses the compatibility path directly.
On ordinary native Linux, the compatibility path uses an available NVIDIA CDI spec before falling back to Docker --gpus all or the NVIDIA runtime.
On Docker Desktop WSL, the compatibility path skips CDI and tries Docker --gpus all before the NVIDIA runtime.
On Jetson/Tegra hosts, the compatibility path uses the NVIDIA runtime and adds eligible host group IDs for the supported GPU device nodes.
These include selected /dev/nvmap, /dev/nvhost-*, and /dev/nvgpu/igpu0/* nodes plus real /dev/dri/renderD* character devices.
After compatibility recreation starts, onboarding keeps the pre-patch container as a rollback backup until the replacement passes the Ready, GPU, and applicable local-inference checks.
If one of those checks fails before backup removal, onboarding prints failure diagnostics and attempts to restore the pre-patch container.
To commit the replacement, NemoClaw stops it, removes the rollback backup, starts the replacement as the final container lifecycle event, and verifies OpenShell supervisor readiness again.
If that final handoff cannot be confirmed, onboarding exits with the container diagnostics and cleanup guidance instead of reporting success.
If rollback fails, onboarding reports that the pre-patch container was not restored and prints container-cleanup guidance.
GPU-proof diagnostics are captured before rollback and can print that guidance before the final container state is known, so inspect the sandbox and its labeled Docker containers before running a deletion command.
Prerequisites:
- Ensure NVIDIA GPU drivers are installed and working.
- On generic NVIDIA hosts,
nvidia-smimust succeed. - On Jetson/Tegra hosts shipping without
nvidia-smi, the devicetree firmware fallback substitutes.
- On generic NVIDIA hosts,
- NVIDIA Container Toolkit configured for Docker.
When GPU passthrough is enabled and a gateway already exists without it, onboarding first checks whether replacing the CPU-only gateway is safe.
If no other registered sandbox depends on that gateway, or if --recreate-sandbox is recreating the only registered sandbox with the same name, onboarding cleans up the stale gateway and continues.
If other sandboxes depend on the gateway or Docker state is unclear, onboarding exits without cleanup and prints targeted destroy or gateway-removal guidance.
To add GPU to an existing sandbox, rerun with --recreate-sandbox.
Leave NEMOCLAW_DOCKER_GPU_PATCH unset or set it to auto for native-only GPU onboarding on ordinary native Linux.
Set NEMOCLAW_DOCKER_GPU_PATCH=fallback to explicitly opt into one bounded native-to-compatibility retry on ordinary native Linux.
Set NEMOCLAW_DOCKER_GPU_PATCH=0 to require native OpenShell GPU injection on ordinary native Linux or Jetson/Tegra.
Set NEMOCLAW_DOCKER_GPU_PATCH=1 to use only the compatibility path on ordinary native Linux.
Other legacy nonzero values keep that behavior through the v0.0.x release line and will be removed in v0.1.0.
Use NEMOCLAW_DOCKER_GPU_PATCH=0 on Jetson/Tegra only for troubleshooting because it bypasses Tegra device-group propagation and CUDA may not initialize.
Docker Desktop WSL ignores NEMOCLAW_DOCKER_GPU_PATCH=0 because GPU passthrough on that runtime requires the compatibility patch.
Use --no-sandbox-gpu, --no-gpu, or NEMOCLAW_SANDBOX_GPU=0 when you want to disable sandbox GPU passthrough on Docker Desktop WSL.
nemo-deepagents list
List all registered sandboxes with their model, provider, and policy presets.
Pass --json for machine-readable output that includes a schemaVersion, the default sandbox, recovery metadata, and the sandbox inventory.
Each sandbox row reports activeSessionCount as a nonnegative integer when the SSH-session probe is available and null when it is unavailable.
Each sandbox row reports agent as a string in both text and JSON output, never null.
The row reports openclaw when the registry records no agent for the sandbox.
The row reports unknown for a sandbox that nemo-deepagents list recovers from the live OpenShell gateway.
The gateway sandbox list does not expose the agent.
The row does not include the former derived connected boolean.
Sandboxes with an active SSH session are marked with a ● indicator so you can tell at a glance which sandbox you are already connected to in another terminal.
The default sandbox in text and JSON output honors the same environment override order as host-level status and tunnel commands: NEMOCLAW_SANDBOX_NAME, then NEMOCLAW_SANDBOX, then SANDBOX_NAME, then the registry default.
nemo-deepagents use <name>
Promote a registered sandbox to the default.
This is the first-class replacement for hand-editing ~/.nemoclaw/sandboxes.json; it updates the registry through the same atomic, lock-guarded path that nemo-deepagents onboard uses for the initial default.
Subsequent commands and the NEMOCLAW_SANDBOX_NAME resolution order then pick up the new default automatically.
Pass --json to receive a machine-readable result indicating whether the registry was updated, the sandbox was already the default, or the name is unknown.
nemo-deepagents use is a thin selector and never mutates the sandbox itself.
It fails with a non-zero exit and a known-sandbox list when the requested name is not registered, so scripts can branch safely on the outcome.
nemo-deepagents launch <name>
Connect to a sandbox and start its agent in one host-side command.
Use it instead of running nemo-deepagents <name> connect and then typing the agent command inside the sandbox.
launch runs the complete preflight from nemo-deepagents <name> connect when no launch-readiness lease is usable.
That path includes the readiness wait, in-sandbox agent process recovery, and inference-route reconciliation.
A successful complete preflight can publish a credential-free launch-readiness lease with a fixed 24-hour lifetime on Linux.
Lease acceptance and publication are currently Linux-only and require a secure, independently writable OS per-user runtime authority under /run/user/<numeric-uid>.
It never uses caller-provided environment variables to select this authority.
On macOS, launch runs the complete preflight every time and does not publish a launch-readiness lease.
During that lease, another launch still verifies these conditions:
- The owning OpenShell gateway reports the exact sandbox identity in the
ReadyorRunningstate. - The sandbox registry, agent manifest, interactive command, policy intent, and effective parsed OpenShell network policy match the recorded identity.
- The recorded inference selection matches the live route, and
inference.localreturns HTTP 2xx from its semantic probe when inference is configured. This is stricter than the HTTP 200–499 reachability diagnostic used by ordinaryconnect. - The agent runtime and its required host-side forwards pass their semantic health checks.
Hermes and LangChain Deep Agents Code retain their existing session setup on the lease-accepted path.
After these checks pass, launch can skip duplicate recovery, readiness polling, and inference-route repair.
The lease does not replace a health check or authorize repair.
For missing, expired, malformed, inaccessible, mismatched, or unhealthy evidence, NemoClaw fences any prior acceptable evidence before it runs the complete preflight.
Ordinary launch continues only when NemoClaw proves that no old authority or evidence can exist, or durably rotates the runtime epoch.
If an old epoch might exist and cannot be durably rotated, launch stops before complete preflight or recovery.
Its redacted guidance asks you to repair the current user’s secure OS runtime authority and NemoClaw state permissions, then retry.
A failed live check never becomes a successful launch because a lease exists.
Immediately before the first mutation in the complete preflight, the producer revalidates its sandbox-global runtime epoch while holding the sandbox lifecycle lock followed by the owning gateway lock. It holds both locks through all mutations in the complete preflight, final state capture, and publication. If another producer has replaced the epoch, the stale producer makes no changes and re-inspects the newer lease.
The 24-hour lifetime does not extend when you launch repeatedly.
Exiting the agent with /exit does not revoke the lease.
If state changes before expiry, NemoClaw fences the old evidence and runs the complete preflight.
A successful preflight in that interval keeps the original start and expiry time.
After expiry, a successful complete preflight starts a new 24-hour lease only when publication succeeds.
If unsafe or malformed authority history makes the prior lease timeline untrustworthy, NemoClaw durably invalidates the old epoch and starts one conservative 24-hour quarantine.
Both wall time and monotonic uptime must span the full quarantine, and publication remains disabled during it.
Repeated attempts do not extend the quarantine.
After it elapses, the next successful complete preflight can publish a new fixed 24-hour lease.
You do not create or refresh this lease manually, and launch has no lease-control flags.
After lease validation or the automatic fallback that runs the complete preflight, launch starts the sandbox’s agent in your terminal instead of opening a sandbox shell.
The agent command comes from the sandbox’s agent manifest.
If the sandbox registry names a non-OpenClaw agent without a readable local agent manifest, launch exits before starting an in-sandbox command.
The sandbox name is required, and the command takes no flags.
The sandbox must already exist in the local NemoClaw state.
If it is not registered locally, launch exits before it runs an OpenShell command or readiness recovery and reports that the sandbox is not registered in the local NemoClaw state.
When the agent exits, you return to the host shell.
launch returns the agent’s exit code.
When you want a shell inside the sandbox rather than an agent session, use nemo-deepagents <name> connect.
nemo-deepagents deploy
The nemo-deepagents deploy command is deprecated.
Prefer provisioning the remote host separately, then running the standard NemoClaw installer and nemo-deepagents onboard on that host.
Deploy NemoClaw to a remote GPU instance through Brev.
This command remains as a compatibility wrapper for the older Brev-specific bootstrap flow.
The Brev instance name is the positional argument.
The sandbox name comes from NEMOCLAW_SANDBOX_NAME and defaults to my-assistant; invalid sandbox names fail before Brev provisioning starts.
nemo-deepagents <name> connect
Connect to a sandbox by name.
Bare nemo-deepagents connect (no sandbox name) connects to the registry default.
NemoClaw uses the stored default when it names a non-pending registered sandbox, then falls back to the first non-pending registration.
If only pending registrations remain, the command exits non-zero and tells you to wait for onboarding or remove the incomplete sandbox.
If the registry remains empty after recovery, it tells you to run nemo-deepagents onboard.
A registered sandbox literally named connect keeps the name-first reading.
If the sandbox is not yet in the Ready phase, connect polls openshell sandbox list every few seconds and prints the current phase. This gives you progress output right after onboarding, when the 2.4 GB image is still pulling, instead of a silent hang.
Control the wait budget with NEMOCLAW_CONNECT_TIMEOUT in integer seconds. An interactive connection defaults to 120 seconds, while --probe-only and nemo-deepagents <name> start default to 300 seconds so a scripted health check can wait through a cold sandbox start. When the deadline expires, connect exits non-zero with the last-seen phase.
On a TTY, a one-shot hint prints before dropping into the sandbox shell.
The hint is agent-aware. It names the correct TUI command for the sandbox’s agent and reminds you to use /exit to leave the chat before exit returns you to the host shell.
Set NEMOCLAW_NO_CONNECT_HINT=1 to suppress the hint in scripted workflows.
If the sandbox is running an outdated agent version, a non-blocking warning prints before connecting with a nemo-deepagents <name> rebuild hint.
If another terminal is already connected to the sandbox, connect prints a note with the number of existing sessions before proceeding. Multiple concurrent sessions are allowed.
Without --probe-only, connect does not pull a model itself, but it does inspect managed-vLLM install variables such as NEMOCLAW_VLLM_MODEL and NEMOCLAW_VLLM_EXTRA_ARGS_JSON if you exported them in the same shell.
An unknown model slug, malformed extra-args JSON, or a gated model (for example deepseek-r1-distill-70b) with no HF_TOKEN or HUGGING_FACE_HUB_TOKEN exits non-zero with the same error the installer would emit, before any sandbox readiness probe or SSH attach.
Unset the managed-vLLM variable, or fix the value, before retrying a regular connection.
connect --probe-only skips this install preflight so stale managed-vLLM variables cannot block recovery.
Before reading or changing the live OpenShell gateway inference route, connect verifies the shared provider and sandbox metadata.
When the live route differs and the metadata is compatible, connect warns and re-points the route to the target sandbox’s recorded provider and model.
Refer to Use Shared Gateway Routes for provider-global identity, route drift, and hard-error recovery.
Use nemo-deepagents inference set --provider <provider> --model <model> to make an intentional compatible route change outside the connect flow.
Before it opens SSH, connect probes https://inference.local/v1/models from inside the sandbox with the selected agent’s trusted CA and proxy context.
HTTP 200 through 499 confirms that the route is reachable.
When the probe returns a recognized broken result, connect attempts DNS or route repair and verifies the route again.
When the initial probe cannot return a trusted result, connect fails closed before health-driven repair and before opening SSH.
It prints a bounded, redacted last-probe detail and points you to nemo-deepagents <name> doctor.
If the sandbox is registered locally but missing from a healthy gateway, connect preserves the registry entry and points you to rebuild --yes, onboard, or destroy instead of deleting the metadata needed for recovery.
After a host reboot, the OpenShell gateway rotates its SSH host keys.
connect detects the resulting identity drift, prunes stale openshell-* entries from ~/.ssh/known_hosts, and retries automatically.
You no longer need to re-run nemo-deepagents onboard after a reboot in this case.
On Linux, the --probe-only flag is the infrastructure producer for launch-readiness evidence.
It validates a usable lease and exits without duplicate recovery.
Otherwise, it fences prior evidence, waits for the sandbox, verifies or repairs its in-sandbox agent process and host-side forwards, and publishes evidence only after every probe succeeds.
It rechecks the sandbox on its recorded OpenShell gateway after the readiness wait and never restarts the shared host gateway.
If an old runtime epoch might exist and cannot be durably rotated, the command exits nonzero before complete preflight or recovery and gives redacted repair guidance.
A securely absent runtime authority and receipt let ordinary launch run the complete preflight without optimization if new authority creation fails, but on Linux connect --probe-only still exits nonzero because it could not publish launch-readiness evidence.
A runtime failure and, on Linux, a failure to publish evidence for an otherwise healthy runtime also exit nonzero with different diagnostics.
Infrastructure must run the command as the same final numeric user that later runs launch.
Run it only after the final durable home and state volume is mounted and after policy and network provisioning is complete.
On Linux, that user also needs a secure, independently writable OS per-user runtime authority under /run/user/<numeric-uid>.
Do not redirect this authority with caller environment variables.
Do not use a graphical or login-session identifier as the deployment ordering boundary.
On macOS, connect --probe-only runs the complete preflight, including recovery and probes.
After a successful probe and recovery, it prints a note that launch-readiness evidence is unavailable on this platform and exits zero.
The next launch runs the complete preflight.
On Linux, the publication-failure diagnostic is redacted and does not print filesystem paths or environment values.
Run it for health checks and scripted readiness probes; users continue to run only nemo-deepagents launch <name>.
Use nemo-deepagents launch <name> when you want launch-readiness validation, an automatic fallback that runs the complete preflight, and then the agent instead of a sandbox shell.
nemo-deepagents <name> exec
Run a single command non-interactively in a running sandbox via the OpenShell exec endpoint.
The command runs as the sandbox user with HOME=/sandbox, so in-sandbox tooling resolves NemoClaw-provisioned config the same way it does for connect and openshell sandbox connect.
This is the supported substitute for docker exec on the sandbox container; raw docker exec runs as root and lands on HOME=/root, where the selected agent config is not present.
For a registered sandbox, NemoClaw selects its recorded owning OpenShell gateway before the workdir probe and command dispatch.
If gateway selection fails, exec stops without running the sandbox command.
Everything after -- is forwarded verbatim to the sandbox command, including flags the inner command needs.
By default, NemoClaw inherits caller stdin only when it is a terminal.
Non-terminal or unavailable stdin is closed so SSH, CI, and other one-shot commands cannot wait on an inherited pipe.
Pass --stdin to forward an intentional pipe, or --no-stdin to close terminal stdin explicitly.
OpenShell preserves line endings and quote characters inside each command argument, so inline scripts and heredocs can be passed as one argument after --.
For example, a shell variable keeps the multi-line script in one argv element:
NUL bytes are still rejected in command arguments.
Line breaks are accepted only in command argv: --workdir remains single-line, and NemoClaw does not expose OpenShell request-environment injection on this command.
nemo-deepagents <name> agent
For Deep Agents sandboxes, agent forwards to the manifest-declared terminal command.
Bare invocations run dcode, and --help runs dcode --help.
Use dcode -n for explicit headless automation when you are already connected to the sandbox, or use nemo-deepagents <name> agent -n "<task>" from the host.
Add --json to either form for one managed, versioned JSON envelope on stdout.
The host wrapper forwards the flag to dcode.
For the schema, status and exit behavior, and 1 MiB output limit, refer to Run Deep Agents Code.
The host wrapper keeps HOME=/sandbox, the managed proxy environment, and the manifest-declared Deep Agents config path aligned with connect.
Interactive nemo-deepagents <name> agent launches the same terminal TUI as dcode.
Headless nemo-deepagents <name> agent -n "<task>" uses the managed headless boundary, where non-shell tools can auto-run without the interactive approval UI.
Advanced Sandbox Maintenance Commands
The following commands are available for targeted host-side maintenance, but they are not part of the top-level public command list.
nemo-deepagents <name> config get
Read the sanitized agent configuration from a sandbox.
The output removes credential-bearing sections before printing.
Use --key to read one dotpath and --format to choose JSON or YAML output.
nemo-deepagents <name> config set
For Deep Agents sandboxes, config set is unavailable because the dcode configuration is baked into the sandbox image at build time.
Run nemo-deepagents onboard --agent dcode --name <sandbox-name> --fresh when you need to change it.
Use nemo-deepagents <name> config get to read the current values.
nemo-deepagents <name> shields
Manage the sandbox config lockdown posture from the host.
Use shields status to inspect the current state, shields up to lock the sandbox config and restore the captured restrictive policy, and shields down to temporarily unlock the config for maintenance.
If OpenShell rejects the permissive policy before it is applied, shields down returns an error and keeps the sandbox in the Shields up state.
The command clears the provisional Shields down record and timer, and shields status remains UP.
If that record cannot be cleared and NemoClaw writes the rejection marker, shields status derives UP from that marker.
The auto-restore timer and transition remain the recovery authority.
If the rejection marker also cannot be written, shields status reports the incomplete transition as an error.
If a config path is unsafe, for example a symlink at the Hermes config.yaml path, shields down refuses that path before it weakens policy, writes a provisional Shields down record, or starts a timer.
The command returns an error and shields status remains UP.
If an unsafe path appears after the preflight and a provisional Shields down record already exists, the command restores the restrictive policy when it can but keeps the Shields down record until config protection is positively re-verified. This fail-closed behavior also applies when unlock fails after a partial mutation, and requires manual intervention if re-lock cannot be confirmed.
If shields up reports that the config remains unlocked or drifted, confirm that the sandbox is running and ready, then retry nemo-deepagents <name> shields up.
If the retry still fails, rebuild a known-good baseline with nemo-deepagents <name> rebuild --yes.
A CRITICAL Deep Agents config-lock failure is not an ordinary unlocked or drifted result.
The retry and rebuild guidance above does not apply to a CRITICAL Deep Agents config-lock diagnostic.
Do not retry shields up or attempt an in-sandbox repair.
Follow Deep Agents Config Lock Failure Recovery to restore a trusted snapshot or recreate the sandbox before retrying.
Host-side config and inference writes, snapshot mutation, sandbox destruction, and shields transitions serialize per sandbox.
Before shields down opens a new window, NemoClaw must revoke any stale auto-restore timer authority.
If marker cleanup fails, the command reports Cannot revoke stale auto-restore timer authority and stops before policy capture, state writes, config unlock, replacement-timer startup, or audit writes.
The sandbox retains its existing configuration and policy posture, and the stale timer authority remains.
Resolve the reported timer-marker error on the trusted host, then retry shields down.
When a timed shields-down window reaches its deadline, auto-restore closes the per-sandbox lifecycle deadline gate.
The gate blocks new mutations and waits for the recorded live owner to release its exact lock generation before auto-restore restores lockdown.
NemoClaw does not signal that process because portable process inspection cannot prove that all descendants stopped.
An interactive command can take over an expired timer.
Interactive recovery has separate transition-takeover and restoration phases.
Each phase makes up to 7 attempts and waits 5 seconds between failures, for up to 30 seconds of retry delay per phase.
Detached recovery uses one 7-attempt budget across deadline setup, main-generation publication, and restoration.
The deadline gate remains closed during those attempts.
If restoration cannot commit, NemoClaw attempts to record durable containment.
If that containment commit also fails, NemoClaw retains any exact lifecycle and deadline gates it already owns.
A state-directory failure that prevented gate publication also prevents normal mutation-lock acquisition.
Correct the reported state-directory write failure, then run nemo-deepagents <name> shields status to resume recovery or receive exact-generation recovery guidance.
When recovery cannot complete, an interactive command returns an error, or the detached timer exits with a failure status.
NemoClaw also records durable containment when an owner exits before it can prove that the owner’s descendants stopped, or when ownership becomes ambiguous.
Durable containment, retained exact gates, or the fail-closed state-directory error blocks new mutations until you complete exact-generation operator recovery.
Stop all NemoClaw processes for the sandbox, then follow the paths, identities, tokens, and removal order in the reported error.
Verify each recorded generation is unchanged, remove only the exact stale generations first, and remove the exact containment generation last.
Before a manual Shields transition replaces a policy, NemoClaw requires exact agreement among the sandbox registry, generated-policy record, and live gateway policy.
shields down carries the proven managed MCP policy entries into the relaxed policy.
Restoration removes snapshot-time managed MCP entries before it overlays current exact entries.
If exact agreement is absent, a manual Shields transition refuses the replacement policy.
At an expired deadline, auto-restore omits unproven managed MCP policy entries, restores lockdown, and records the omission count in its audit entry.
An MCP server removed during the shields-down window stays removed.
A surviving server keeps its recorded endpoint and address pins while its policy ownership remains exact.
nemo-deepagents <name> stop
Stop the sandbox’s local runtime container while preserving all of its state.
Workspace files, credentials, network policies, the registry entry, and the OpenShell sandbox record stay in place.
Use this to free CPU, memory, and GPU resources without destroying the sandbox; use nemo-deepagents <name> destroy when you want to delete it instead.
For OpenClaw-managed gateways, the command first asks the in-sandbox gateway to shut down its channels gracefully; non-Portable agent-managed gateways (for example Hermes) are supervised inside the sandbox and shut down with the container’s stop signal. Then the container stops; a container stuck in a crash loop is stopped the same way, which also disarms its restart policy. The shared host gateway, tunnel services, and any local NIM inference container serve other sandboxes and keep running. Stopping an already-stopped sandbox succeeds without changes. The command is available only when NemoClaw holds local-container authority. Portable profiles use receipt-owned Podman authority; non-Portable local-container paths use the default Docker driver or the vm driver. Remote drivers such as kubernetes are unavailable, and an unreachable selected runtime produces an outage report instead of a guessed container state.
nemo-deepagents <name> start
Restart a sandbox container that was stopped with nemo-deepagents <name> stop or by a host reboot, then repair the in-sandbox gateway and host-side forwards the same way nemo-deepagents <name> recover does.
Starting an already-running sandbox skips the container start and still runs the gateway and forward health checks.
A paused container is unpaused.
If the container was removed entirely, start fails and points you to nemo-deepagents <name> rebuild.
Before it verifies the managed terminal runtime, start waits for OpenShell to report the sandbox in the Ready or Running state, using the same 300-second budget and NEMOCLAW_CONNECT_TIMEOUT override as connect --probe-only.
When that deadline expires, start keeps the existing container, exits non-zero, and prints the NEMOCLAW_CONNECT_TIMEOUT value to use on the next run.
After the gateway and forward checks pass, start sends one inference request through https://inference.local using the sandbox’s recorded provider and model.
A gateway that answers the /v1/models probe can still reject an inference request or return an invalid result, so the command exits non-zero in either case.
It prints the probe result, including the HTTP status when the route returned one, and points you to the sandbox doctor command.
Each run sends one 16-token request through the stored provider credential, so start waits up to 30 seconds for it and consumes provider tokens on a hosted route.
When the sandbox records no provider or no model, start skips the request and exits 0.
doctor still classifies an HTTP 401 or 403 route response as reachable, so correct the provider credential when start reports one of those statuses.
nemo-deepagents <name> status
Show sandbox-scoped status, health, and inference configuration for one registered sandbox.
Use this form when you care about a specific sandbox’s live OpenShell state, agent runtime, inference health, GPU proof, permissions, and recovery hints.
Do not pass a sandbox name to nemo-deepagents status; that command is the global all-sandbox/service overview.
NemoClaw resolves the sandbox’s recorded owning OpenShell gateway before querying live state.
If another gateway is active, it selects the owner and queries again instead of trusting a result from the sibling gateway.
For a compatible-endpoint route that uses openai-completions, the text output prints Reasoning effort as low, medium, high, or endpoint-default.
The line is omitted for another provider or API family.
Pass --json to emit a structured per-sandbox report instead of the text renderer.
The JSON output includes at least schemaVersion, name, found, agent, agentDisplayName, agentRuntime, dcodeAutoApprovalMode, model, provider, recordedRoute, liveRoute, routeDrift, phase, gatewayState, inferenceHealth, rpcIssue, hostGpuDetected, sandboxGpuEnabled, sandboxGpuMode, sandboxGpuDevice, openshellDriver, openshellVersion, policies, baselineExclusions, baselineExclusionStates, baselineExclusionTransition, failureLayer, terminalRuntimeHealth, servingProcessHealth, and dockerPaused.
baselineExclusions is an array of exact baseline keys recorded for durable replay and is empty when the sandbox has none.
baselineExclusionStates reports each recorded key with its current verification state.
The excluded state means the reviewed entry still matches the active agent baseline and the key is absent from the live OpenShell policy.
Other states identify agent drift, changed or removed baseline content, an unreadable baseline or live policy, or a live policy that contains the excluded key.
baselineExclusionTransition is null when policy state is settled; otherwise it identifies the interrupted exclude or restore key that must be reconciled before sandbox creation or recreation, rebuild, or cross-sandbox snapshot cloning.
The schema-version 1 model and provider fields keep their established live-route meaning when the gateway route is readable.
Use recordedRoute for the sandbox’s durable provider and model and liveRoute for the gateway-global route.
When the live shared route differs, text output prints both routes and JSON output sets routeDrift.live, routeDrift.recorded, and routeDrift.canConnect.
When routeDrift.canConnect is false, connect cannot safely restore the recorded route because provider-global identity differs or required route or gateway metadata is incomplete.
Refer to Use Shared Gateway Routes for the route-sharing workflow.
openshellDriver and openshellVersion are always strings (falling back to "unknown" when the registry has no value), so consumers can rely on typeof checks.
agent is always a string and reports openclaw when the registry records no agent for the sandbox.
failureLayer is null when no preflight failure was detected and otherwise one of docker_unreachable, sandbox_container_stopped, or sandbox_dashboard_port_conflict; when set, inferenceHealth is suppressed to null so automation does not see a stale remote-provider healthy status during a local outage.
inferenceHealth.ok reports whether the inference route returned a structurally valid result for one request sent from inside the sandbox.
The result must match Chat Completions, Responses, or Anthropic Messages for the selected route.
An empty body, malformed JSON, provider-error envelope, or wrong response shape reports unhealthy, even with a 2xx status.
The probe captures at most 64 KiB and does not include the response body in diagnostics.
The route probe treats any final HTTP status from 200 through 499 as reachable, so a route with an invalidated provider credential answers HTTP 401 while the route is up.
The request uses the live gateway route’s provider and model, and falls back to the recorded values when the live route is unreadable.
When the live provider matches the recorded provider, the request uses the sandbox’s recorded API family, even when only the model differs.
This includes openai-responses.
When the live provider differs, NemoClaw does not carry the recorded API family to the live provider.
An ordinary run sends one 16-token request through the stored provider credential, with a 30-second timeout, and consumes provider tokens on a hosted route.
When the same status run recovers a managed gateway, it retries the route and inference request together up to three total attempts, with a two-second delay between failed attempts.
Each attempt can consume another 16 tokens on a hosted route.
When NemoClaw sends an inference request, inferenceHealth.subprobes reports the route probe result as the route reachability hop, so a failing verdict still shows that the route itself answered.
inferenceHealth.failureLabel reports why the inference request failed:
unauthorizedwhen the route rejected it with HTTP401or403.unhealthywhen the route returned another failing HTTP status or an invalid 2xx response body.unreachablewhen the request returned no HTTP status, including a probe that could not run.
A host-side upstream probe under inferenceHealth.subprobes stays a diagnostic and does not change inferenceHealth.ok, because the sandbox route is the one the agent uses.
When the route probe failed, or the sandbox records no provider or no model, NemoClaw skips the inference request and inferenceHealth reports the route probe result alone.
dockerPaused is true when NemoClaw detects that the Docker-driver sandbox container is paused.
In that case, text output keeps OpenShell’s authoritative phase but prints a docker unpause <container> recovery hint instead of sending you directly to rebuild.
For terminal runtime sandboxes, the command also checks cgroup OOM kill counters.
If the counter records an OOM kill, text output prints Runtime health: degraded (... OOM kill recorded) and points you to nemo-deepagents <name> rebuild; JSON output reports terminalRuntimeHealth.kind: "degraded" with the OOM kill count and source counter path.
For a present gateway runtime, text output prints Serving process (<agent> gateway): not checked, and JSON output reports servingProcessHealth: { "checked": false }.
The existing inference probes run in a fresh sandbox command, so they do not attest that the long-running gateway process has equivalent inference access.
NemoClaw does not probe the serving process yet.
For terminal runtimes, servingProcessHealth is null and the text output omits this line because there is no long-running gateway process.
The command exits non-zero when the sandbox is missing locally, the gateway state is not present, the gateway reports a schema/protobuf mismatch (mirrored as rpcIssue), failureLayer is non-null, the authoritative in-sandbox inference route fails or cannot be probed, or a terminal runtime sandbox reports a recorded OOM kill.
When the canonical text command targets an unregistered name, it reports that the sandbox is not registered and tells you to run nemo-deepagents list.
The alias form nemo-deepagents <name> status --json requires the sandbox to be registered locally; the canonical form nemo-deepagents sandbox status <name> --json is the one to use from automation that may run against an unknown sandbox name, since it still emits a JSON document with found: false instead of a text error.
For a sandbox that owns managed llama.cpp, text output also reports the recipe ID, model digest, image reference, https://inference.local/v1 endpoint, and lifecycle state.
It does not print the managed API key or its fingerprint.
The lifecycle state is one of these values:
The managed llama.cpp check forces a nonzero exit for absent, conflict, or unknown.
Other sandbox and inference checks can also make the command fail.
Rerun the same NEMOCLAW_PROVIDER=install-llama-cpp and NEMOCLAW_LLAMACPP_RECIPE onboarding selection to recover a stopped or interrupted runtime.
Inspect and correct an identity conflict before retrying.
For a Deep Agents sandbox, text output includes DCode auto-approval capability: disabled or DCode auto-approval capability: thread-opt-in.
JSON output reports the same configured value in dcodeAutoApprovalMode.
This value does not attest that auto-approval is active in any live TUI thread.
The command probes https://inference.local/v1/models from inside the sandbox, and when that probe reports the route reachable it sends one inference request over the same route.
That inference request is the authoritative inference health check, and both checks exercise the route that agent traffic uses.
The main Inference line reports one of these states:
An authentication response on the route probe alone confirms that the route is reachable, not that provider credentials are valid.
nemo-deepagents <name> doctor sends no inference request, so it reports an HTTP 401 or 403 route response as reachable and exits 0 where status reports unauthorized.
The command can also print direct host-side provider checks such as Inference (upstream) and provider-specific subprobes.
For supported remote providers, this diagnostic sends an authenticated request to the configured model and accepts only a recognized Chat Completions, streaming Chat Completions, or Anthropic Messages response.
It uses a 3-second connection timeout, a 5-second total timeout, and an 8-token output limit.
If the request reaches the time limit, NemoClaw reports the provider as not probed and leaves model health unverified instead of reporting it as unhealthy.
These checks are diagnostic only and do not override the authoritative inference.local result or determine the command exit status.
The Inference (upstream) check authenticates with the host credential that NemoClaw resolves for the provider, such as NVIDIA_INFERENCE_API_KEY.
The gateway stores the provider credential that the sandbox route uses.
The CLI cannot read the stored value back, so the two credentials can hold different secrets.
When the inference.local route has already served the inference request, an unauthorized result on Inference (upstream) describes the host credential.
NemoClaw then reports that check as not probed and names both credential sources.
An Inference (upstream) check that fails for another reason, such as unreachable, still reports its own state.
Local backend and auth proxy checks, such as Inference (auth proxy), always report their own state and their own repair step.
nemo-deepagents <name> doctor sends no inference request, so it always reports the Inference (upstream) state that it measured.
Local providers add host-side backend diagnostics.
For Local Ollama, the command can also print an Inference (auth proxy) diagnostic when a proxy token is available.
Use these diagnostics to identify a failing auxiliary hop after checking the main Inference line.
For cloud-only providers, the output omits the NIM status line unless a NIM container is registered or an unexpected NIM container is running.
When the sandbox’s recorded driver is docker and the host Docker daemon is not reachable, the command prints the docker_unreachable failure layer with the message Docker daemon is not reachable. as the first line of stdout, suppresses the host-side Inference probe (which otherwise hits the remote provider directly and is misleading when the local stack is down), and exits with a non-zero status.
When the host Docker daemon is reachable but the per-sandbox container is stopped, the initial preflight records the sandbox_container_stopped failure layer and suppresses the host-side Inference probe.
If the owning OpenShell gateway is healthy but no longer lists the registered Docker-driver sandbox, status attempts post-reboot recovery from the labeled container.
It waits for Docker readiness, restores the in-sandbox gateway and host forwards, and refreshes preflight before probing inference.
A successful recovery clears the stale stopped-container failure.
When status finds the sandbox but cannot prove its agent delivery chain, it exits non-zero and reports the sandbox_recovery_failed state.
Address the reported recovery layer, then run the displayed nemo-deepagents <sandbox-name> recover command.
If the sandbox or gateway cannot be verified, the command exits non-zero instead of reporting healthy inference from stale registry state.
When a locally registered sandbox is missing from the live gateway, status preserves the registry entry so the suggested rebuild --yes recovery can still find the sandbox metadata.
When sandbox GPU passthrough is enabled, the Sandbox GPU line includes the last CUDA usability proof state.
It reports (CUDA verified), (CUDA unverified), or (last CUDA proof failed: <label>) so automation and operators can distinguish configured GPU passthrough from proven CUDA access.
Failed proofs include remediation guidance for the detected platform.
An SSH sessions line reports how many active SSH sessions the sandbox has, or none; the line is omitted when the session probe is unavailable.
The Policy section displays the live enforced policy (fetched via openshell policy get --full), which reflects presets added or removed after sandbox creation.
When OpenShell reports an active policy version, the displayed YAML version line uses that active version instead of the static schema version.
If the sandbox is running an outdated agent version, the output includes an Update line with the available version and a nemo-deepagents <name> rebuild hint.
Checking the Deep Agents version
Refer to Update Sandboxes for the Deep Agents Code version pin and rebuild policy.
nemo-deepagents <name> status prints the running Deep Agents Code version on the Agent line:
Expected output:
If the sandbox is running an older Deep Agents Code version than this NemoClaw release expects, status and connect add an Update line pointing at nemo-deepagents <name> rebuild to pick up the newer version.
The rebuild reuses the existing sandbox name and preserved manifest-defined state, so skills, app state, and managed config carry over while credentials stay in host-side OpenShell state.
nemo-deepagents <name> doctor
Run a focused health check for one sandbox and the host services it depends on. The command checks the local CLI build, Docker daemon, OpenShell CLI, NemoClaw gateway container, gateway port mapping, live sandbox state, inference route, configured-provider model invocation, Ollama reachability, and the cloudflared tunnel state.
doctor also checks whether the sandbox registry contains the metadata required for snapshot, rebuild, upgrade, recovery, and reboot.
When lifecycle metadata is incomplete, the report names the missing or invalid fields and affected operations without printing stored values.
Dashboard metadata is required only for agents that manage a dashboard.
If the registered gateway binding is invalid, doctor reports a failed gateway check and does not select, probe, or recover a gateway from that binding.
For inference health, doctor treats the probe to https://inference.local/v1/models from inside the sandbox as authoritative.
HTTP responses from 200 through 499, including 401 and 403, pass this check.
HTTP 500 through 599, interim 100 through 199, transport failures with status 000, invalid status values, and an unavailable authoritative probe fail the check.
Direct provider and upstream probes use the same authenticated model-invocation checks as status and remain diagnostic only, so their failure does not fail doctor when the authoritative in-sandbox route is reachable.
For gateway runtimes, doctor also reports an informational Serving process: not checked result because its fresh sandbox probes do not attest the long-running gateway process.
This result does not fail the readiness check.
Terminal runtimes omit it because they have no long-running gateway process.
The Inference Route check warns when either the provider or model is unknown.
After the gateway is healthy, run nemo-deepagents <name> status to refresh the route information.
For each recorded baseline exclusion, doctor compares the approval with the active agent baseline and verifies that the excluded key is absent from the live OpenShell policy.
An unreadable live policy produces a warning because NemoClaw cannot verify enforcement.
A live policy that contains the excluded key fails the check and requires policy repair before you rely on the exclusion.
Warnings do not make the command fail.
Failed checks, including a failed or unavailable authoritative inference route, exit non-zero so scripts can use doctor as a readiness gate.
Use --json for machine-readable output.
For a compatible-endpoint route that uses openai-completions, the JSON report includes an informational Inference check labeled Reasoning effort.
The check reports low, medium, high, or endpoint-default and never includes credentials.
Because the check has info status, it does not change the command’s exit status.
For a sandbox that owns managed llama.cpp, doctor adds secret-free identity and runtime checks.
The runtime check passes only when the exact container is running.
It warns for preparing or stopped, and it fails for absent, conflict, or unknown.
The recovery hint tells you to rerun onboarding for the same sandbox so NemoClaw can use the persisted receipt and create journal.
nemo-deepagents <name> exec
Run a command non-interactively inside a running sandbox through the OpenShell exec endpoint.
The command runs as the sandbox user with HOME=/sandbox.
Use -- to separate exec options from the command you want to run inside the sandbox.
By default, NemoClaw inherits caller stdin only when it is a terminal.
Non-terminal or unavailable stdin is closed so SSH, CI, and other one-shot commands cannot wait on an inherited pipe.
Pass --stdin to forward an intentional pipe, or --no-stdin to close terminal stdin explicitly.
OpenShell preserves line endings and quote characters inside each command argument, so inline scripts and heredocs can be passed as one argument after --.
For example, a shell variable keeps the multi-line script in one argv element:
NUL bytes are still rejected in command arguments.
Line breaks are accepted only in command argv: --workdir remains single-line, and NemoClaw does not expose OpenShell request-environment injection on this command.
nemo-deepagents <name> logs
View sandbox logs.
Use --follow to stream output in real time.
Use --tail <lines> or -n <lines> to limit the number of returned lines.
Use --since <duration> to show recent logs only, such as 5m, 1h, or 30s.
The command reads both agent gateway output and OpenShell audit events, so policy denials appear alongside the gateway log stream.
If one log source is unavailable, NemoClaw prints a warning and keeps reading the remaining source.
NemoClaw’s --tail <lines> flag is a line-count flag; the lower-level openshell logs --tail flag means follow live output, so use openshell logs <sandbox> -n <lines> when running OpenShell directly for a fixed line count.
nemo-deepagents <name> dashboard-url
dashboard-url is not applicable to Deep Agents sandboxes because the managed harness is a terminal runtime without a dashboard port.
Use nemo-deepagents launch <name> to start dcode.
Use nemo-deepagents <name> connect instead when you want a sandbox shell.
nemo-deepagents <name> gateway-token
gateway-token is not applicable to Deep Agents sandboxes because there is no OpenClaw gateway token.
Model traffic uses the OpenShell-managed inference.local route configured by NemoClaw.
nemo-deepagents <name> destroy
Stop managed local inference resources, remove the host-side Docker image built during onboard, and delete the sandbox.
This removes the sandbox from the registry.
For Ollama-backed sandboxes, destroy also asks Ollama to unload currently loaded models and clears stale auth proxy state on a best-effort basis.
For Model Router sandboxes, destroy keeps the process and recovery identity when another sandbox uses the port or when session, process, or absence checks are inconclusive.
It also preserves a replacement onboarding session when the captured session identity changed.
If the captured session uses the destroyed sandbox name with another router port, destroy clears only the sandbox association and preserves that router’s recovery identity.
For lock order, same-port peer handling, and cleanup checks, refer to Set Up Model Router.
If destroy warns that it could not identify or stop a listener for the deleted sandbox:
- Inspect the current listener process immediately before you stop anything.
- Stop it only if its command line identifies the Model Router on the named port.
- Do not stop the router recorded by a preserved session for another port.
- Do not stop a previously reported process ID if its command line no longer matches.
This command attempts to wipe the manifest-defined agent state while its persistent volume is mounted, then removes the sandbox.
OpenShell can retain the per-name persistent volume after sandbox deletion.
If the wipe cannot complete, onboarding with the same name can resurface old files.
Do not rely on a retained volume as a backup.
Back up your workspace first with nemo-deepagents <name> snapshot create or refer to Create and Restore Snapshots.
If you want to upgrade the sandbox while preserving state, use nemo-deepagents <name> rebuild instead.
If another terminal has an active SSH session to the sandbox, destroy prints an active-session warning and requires a second confirmation before it proceeds.
Pass --yes, -y, or --force, or set NEMOCLAW_NON_INTERACTIVE=1, to authorize deletion without prompting in scripted workflows.
Before changing a Docker-backed sandbox, NemoClaw inspects every container with the requested openshell.ai/sandbox-name label.
The command continues when Docker returns no matching containers.
For one matching container, the command continues only when all these labels have the required values:
openshell.ai/managed-by=openshell- A nonempty
openshell.ai/sandbox-workspace - A nonempty
openshell.ai/sandbox-id
If the initial inspection cannot complete, more than one container matches, a matching container has conflicting or incomplete labels, or Docker returns malformed identity data, destroy exits before changing sandbox resources.
The identity checks still apply with --force, --yes, or NEMOCLAW_NON_INTERACTIVE=1; those controls authorize confirmation but do not authorize an unproven container identity.
NemoClaw rechecks the exact identity after read-only preflight, before provider cleanup, and synchronously at the sandbox-deletion boundary.
If a later recheck detects drift or fails, destroy refuses sandbox deletion, restores managed MCP preparation when possible, preserves local ownership state, and reports any earlier cleanup already performed.
If OpenShell reports the sandbox absent after preflight captured one matching Docker container, destroy rechecks and removes only that exact container ID.
If Docker reports another OpenShell-managed container, removal fails, or NemoClaw cannot confirm removal, destroy exits nonzero and preserves the registry entry.
Correct the reported Docker state, then rerun destroy.
If Docker cannot complete the inspection, correct the reported Docker error before you rerun destroy.
For common recovery steps, refer to Docker is not running and Docker permission denied on Linux.
If destroy reports conflicting, incomplete, or malformed identity data, inspect the matching containers:
The labels show what each container claims. They do not prove container ownership.
Do not remove or recreate a container until you verify its purpose, ownership, and data-retention requirements. Removing or recreating a container can discard state that is not stored in a volume.
Resolve a conflict through the workflow that created the conflicting container.
Docker cannot change labels on an existing container.
Rerun the query after you resolve the conflict.
Rerun destroy only when the query returns one complete expected label set that you verified belongs to the target sandbox, or no containers after you independently confirm that the sandbox is absent.
If a shields auto-restore timer is active, destroy holds the same per-sandbox transition through state wipe and deletion.
It restores and verifies lockdown and revokes the active timer before deletion.
It clears the remaining local shields state only after deletion succeeds.
If the pre-delete re-lock fails, the command warns and attempts to destroy the sandbox.
If the destroy operation succeeds, it destroys the sandbox and deletes its unguarded configuration.
If the destroy operation fails, NemoClaw keeps the local shields state and the auto-restore timer.
The timer keeps retrying lockdown and can restore it after the sandbox is reachable again.
NemoClaw records Shields down until a retry verifies lockdown, or until you destroy or rebuild the sandbox.
If deletion fails after hardening, the command keeps the surviving sandbox’s locked shields state instead of cleaning it up as though deletion succeeded.
By default, unattended final-sandbox destroys (--yes, --force, or NEMOCLAW_NON_INTERACTIVE=1) remove the shared NemoClaw gateway on macOS so the host listener is released, while Linux preserves it for reuse.
Pass --cleanup-gateway to force removal, or --no-cleanup-gateway to force preservation.
These flags always override both NEMOCLAW_CLEANUP_GATEWAY and the platform default.
If the pre-delete workspace wipe cannot run, use a different sandbox name for a clean start.
Cleaning up the gateway after the last sandbox also purges the shared cluster volume that retains the per-name persistent volume.
If final gateway cleanup finds a live PID-file process whose command line does not prove it owns the target gateway, destroy exits non-zero after sandbox and registry deletion and skips gateway and volume removal.
NemoClaw preserves the per-gateway PID file and runtime marker so you can inspect the process.
Stop only the listener that matches the target gateway, then rerun destroy to converge cleanup.
When the default-port gateway runs under the packaged OpenShell gateway service, gateway cleanup stops that service before it reaps host processes, so the gateway port is released instead of being rebound by the service manager.
The service is stopped, not disabled or removed, and the next onboarding run starts it again.
On headless Linux, the packaged service can exist while its systemd user manager is unavailable and the gateway runs through the standalone fallback.
For this recognized manager-unavailable failure only, destroy uses the per-gateway PID file when the service is not enabled for automatic activation.
If the recorded PID is live, its command line must match the exact gateway name and port before destroy stops it.
If the recorded process has exited, destroy continues only after it verifies that the gateway port is free.
If a live PID does not prove gateway ownership or the port remains occupied, destroy exits non-zero and preserves the runtime evidence for inspection.
For any other service stop failure, destroy exits non-zero after sandbox and registry deletion, prints the status command for the service, and skips gateway and volume removal.
If the OpenShell gateway is unreachable and the sandbox has no managed MCP ownership state, --force removes only NemoClaw’s local registry entry and local artifacts.
Gateway-side deletion remains unconfirmed, shared host-service and gateway teardown are skipped, and the sandbox and retained volume may still exist if the gateway returns.
Start the gateway with nemo-deepagents <name> status and retry destroy when you need a confirmed deletion.
Managed MCP ownership disables the local-only fallback because exact provider cleanup requires the retained ownership state, and other delete failures remain fatal.
A failed pre-delete re-lock also disables the local-only fallback, because the auto-restore timer is then the only authority that can lock the configuration again after the gateway returns.
After OpenShell confirms deletion of a sandbox that owns managed llama.cpp, destroy revalidates the exact container, internal network, lifecycle journal, and gateway-scoped receipt.
It removes those resources by inspected ID, then removes the API key and managed ownership state.
It preserves the shared ~/.cache/huggingface/ cache.
If exact cleanup fails, destroy preserves the sandbox registry entry and ownership state so you can correct the reported conflict and retry.
nemo-deepagents <name> policy get
Export the sandbox’s round-trippable OpenShell base policy as YAML.
The command runs openshell policy get --base, validates the returned policy, and strips the OpenShell metadata header.
The default output is suitable for review, editing, and later use with openshell policy set.
The command exits non-zero when OpenShell fails, returns an empty response, or returns content that is not valid policy YAML.
Use --raw only to inspect the unparsed OpenShell response, including its metadata header:
Do not pass --raw output to openshell policy set because the metadata header is not part of the policy document.
nemo-deepagents <name> policy add
Add a policy preset to a sandbox. Presets extend the baseline network policy with additional endpoints. Before applying, the command shows which endpoints the preset would open and prompts for confirmation. The scope comes from the exact preset YAML and includes each endpoint’s host, port, access, protocol, TLS, and enforcement settings, allowed methods and paths, and binary allowlist. When a lifecycle operation reapplies a preset, NemoClaw compares it with the live policy and reports whether the preset opens new egress, replaces a drifted entry, or is already effective with no new egress.
To apply a specific preset without the interactive picker, pass its name as a positional argument:
The positional form is required in scripted workflows.
Set NEMOCLAW_NON_INTERACTIVE=1 instead of --yes if you want the same behavior from an environment variable.
Without a preset name, a run with NEMOCLAW_NON_INTERACTIVE=1 reports that non-interactive mode requires a preset name.
A run without a terminal on stdin instead reports that no input is available on stdin.
Both exit non-zero rather than open the picker.
If the preset name is unknown, the command exits non-zero with a clear error.
If a named preset is already applied, the command compares the preset content with the live policy.
When the content matches, the command reports no changes and exits zero.
When the content differs, the command shows the normal preview and asks for confirmation before applying the preset again.
This includes changes to the preset file.
The comparison requires both the preset content and the live policy.
If either cannot be read, the command exits non-zero.
The command also exits non-zero when the name belongs to a custom preset applied with --from-file.
Use --from-file to apply that custom preset again.
Built-in preset choices are scoped to the sandbox’s active agent.
Messaging channel presets appear only when NemoClaw has a matching channel policy for that agent; unavailable channel presets use the standard unknown-preset error before endpoint preview or confirmation.
When a baseline key is durably excluded, NemoClaw reserves that key and refuses built-in, custom, channel, and MCP policy additions that would define it again.
Restore the baseline entry before applying a preset that intentionally owns the same key, or rename a custom preset entry whose key represents different access.
Custom preset files are tracked with the sandbox that applied them.
policy list, policy add, and policy remove compare the local registry and live gateway state using that sandbox-scoped preset metadata, so custom presets do not appear missing just because they are not part of the built-in preset catalog.
Before policy add writes a merged policy, it reads and parses the round-trippable base policy from OpenShell.
If the base policy read returns non-empty output that NemoClaw cannot parse, the command exits non-zero instead of overwriting the live policy with only the new preset.
Fix the gateway or policy read problem, then rerun the command.
For custom presets, the command also reports when the preset reached the gateway but NemoClaw could not record it in the local sandbox registry, because unrecorded custom presets will not appear in policy list or status.
Recover or re-onboard the sandbox, then re-apply the custom preset.
For built-in presets in that same case, the command applies the preset and returns success, because a built-in preset stays discoverable from the gateway.
It warns that policy list will report the preset as active on gateway, missing from local state.
With --from-file or --from-dir, pass a repeatable --trusted-private-host <exact-host-or-ip> option to admit matching RFC1918, carrier-grade network address translation (CGNAT), or IPv6 unique local endpoints.
The option is invalid for built-in presets.
You can supply exact hosts through NEMOCLAW_TRUSTED_PRIVATE_HOSTS instead, and NemoClaw combines the variable with command options.
NemoClaw resolves each matching exact host and adds generated allowed_ips pins to an in-memory copy of the preset.
User-authored allowed_ips remains rejected.
Dry-run output shows the generated pins, and rebuild replays the transformed preset recorded in the sandbox registry without widening it from ambient DNS.
A snapshot alone does not grant private-host authority to a clean target; after a cross-sandbox restore, reapply the source preset with explicit trust.
Use --dry-run to audit a preset before applying it:
Apply a custom preset file when you need to grant access to an endpoint that is not covered by a built-in preset:
For a trusted private endpoint, preview the generated pins before applying them:
For batch workflows, apply all preset files from a directory:
Review every host in custom preset files before applying them. Custom presets bypass the built-in preset review process and can widen sandbox egress.
nemo-deepagents <name> policy list
List available policy presets and show which ones are applied to the sandbox.
The available built-in rows are scoped to the sandbox’s active agent, so unsupported messaging channel policies are not listed for agents without matching channel policy files.
The command cross-references the local registry against the live gateway state (via openshell policy get), so it flags presets that are applied in one place but not the other.
This catches desync caused by external edits to the gateway policy or stale registry entries after a manual rollback.
Preset summaries come only from the YAML preset.description field.
NemoClaw does not render network-policy rule bodies as prose in policy list output.
Recorded baseline exclusions appear in a separate section.
active means the reviewed digest still matches the current baseline, baseline changed — re-review required means the current entry differs, and baseline entry removed — restore to clear means the current release no longer defines the key.
Use status or doctor to additionally verify that the approval belongs to the active agent and the excluded key is absent from the live policy.
repair required — interrupted exclude/restore; rebuild blocked means NemoClaw preserved a durable transaction journal after a crash or persistence failure; rerun the displayed exact policy command to reconcile it before sandbox creation or recreation, rebuild, or cross-sandbox snapshot cloning.
Each active preset is annotated with its provenance so you can tell why it is applied:
[from <tier> tier]— the preset name matches an entry in the sandbox’s current tier definition (see Policy Tiers).[from <agent> agent]— the preset name matches a NemoClaw-managed agent preset and the active agent matches that label.[user-added]— anything else: presets applied later throughpolicy add, presets that match no tier or agent default, or presets that match the opposite agent’s reserved names on a sandbox running the other agent.[source unverified]— the row is active but the local registry and live gateway state disagree. When the gateway cannot be queried, this renders as[source unverified (gateway unreachable)]. The provenance check is suppressed in these trust-degraded states because the source cannot be confirmed against both halves of the sandbox policy view.
Provenance tags are inferred from the sandbox’s current tier and agent metadata at display time and are not persisted per preset.
A preset whose name appears in the sandbox’s current tier YAML is labelled [from <tier> tier] even when an operator added it manually with policy add after onboarding.
Agent-specific preset names are only labelled [from <agent> agent] when the active agent matches that label.
nemo-deepagents <name> policy remove
Remove a previously applied policy preset from a sandbox. The command lists the presets the local registry records together with the presets the live gateway enforces, prompts you to select one, shows the endpoints that would be removed, and asks for confirmation before narrowing egress.
To remove a specific preset non-interactively, pass its name as a positional argument:
Set NEMOCLAW_NON_INTERACTIVE=1 as an alternative to --yes.
Without a preset name, policy remove reports the same two picker errors as policy add and exits non-zero.
If the preset is unknown, or neither the local registry nor the live gateway holds it, the command exits non-zero with a clear error.
A preset the gateway enforces without a local registry record is removable, which is the state policy list reports as active on gateway, missing from local state.
When NemoClaw cannot query the gateway, the command checks the local registry alone; with a preset name it also reports that the gateway could not be queried.
Unchecking a preset in the onboard TUI checkbox also removes it from the sandbox.
nemo-deepagents <name> policy exclude <key>
Persistently exclude one exact entry from the agent baseline policy after previewing the egress and support impact that the change removes.
The preview names the supported features that may stop working.
The command refuses an entry that does not have a reviewed feature-impact disclosure.
The versioned exclusion record is bound to the reviewed baseline content and active agent, then replayed during rebuild.
If the active agent or entry changes, rebuild fails closed until you clear or review the exclusion again.
The command refuses to exclude a key that an applied preset already owns, because removing that live key would also remove the preset’s access.
The critical managed_inference entry cannot currently be excluded pending product direction.
Use --force or --yes for explicit non-interactive acknowledgement, or --dry-run to preview without changing the sandbox.
A run with NEMOCLAW_NON_INTERACTIVE=1, or a run without a terminal on stdin, does not prompt and requires one of those acknowledgement flags.
When a release changes an excluded entry, first run nemo-deepagents <name> policy restore <key> --dry-run to preview the current baseline egress that restoration would allow again.
After you review the output, run nemo-deepagents <name> policy restore <key> --force to allow that egress again and clear the stale exclusion record.
Then preview the current exclusion scope with nemo-deepagents <name> policy exclude <key> --dry-run and reapply it with nemo-deepagents <name> policy exclude <key> --force only if you still accept the support impact.
When a release removes the entry, first run nemo-deepagents <name> policy restore <key> --dry-run to confirm that restoration will clear only the stale exclusion record.
After you review the output, run nemo-deepagents <name> policy restore <key> --force; there is no replacement scope to review or approve.
nemo-deepagents <name> policy restore <key>
Restore a previously excluded entry from the current agent baseline and clear its durable exclusion record.
When the current baseline still defines the entry, --dry-run lists the egress that restoration would allow again.
After you review the output, --force allows that egress again and clears the exclusion record.
When the baseline no longer defines the entry, --dry-run states that restoration will clear only the stale exclusion record.
After you review the output, --force clears that record without changing live egress.
Both paths require explicit acknowledgement unless you use --dry-run; use --force or --yes for non-interactive acknowledgement.
As with policy exclude, a run with NEMOCLAW_NON_INTERACTIVE=1, or a run without a terminal on stdin, does not prompt.
If a restore is interrupted, NemoClaw finalizes it only when the durable exclusion still exactly matches the staged exclusion and the current release baseline still exactly matches the journaled live target.
If either value changed or the current baseline is unreadable, the journal remains in repair required state so you can inspect and re-review the current scope instead of silently accepting a different entry.
The restore command accepts these flags:
nemo-deepagents <name> policy explain
Print a redacted summary of the active policy context for a sandbox so an agent or operator can reason about what is allowed, what is blocked, and how to request a change. The output covers the recorded tier, applied presets and allowed host categories, known unapplied presets, baseline exclusions and their support impact, policy-change commands, and the support boundaries between NemoClaw, OpenShell, and the agent. Raw policy YAML, rule bodies, and credential metadata are deliberately not included.
Pass --json to emit the same context as a structured object for agent consumption:
The context also documents how a failed host or integration attempt should be classified.
The classifications are blocked-by-policy, missing-approval, unsupported, and unknown, so the agent can pick a remediation step instead of surfacing a lower-level network error.
nemo-deepagents <name> hosts-add
Add a host alias to the sandbox pod template.
Use this when a sandbox needs a stable LAN-only name, such as a local SearXNG or internal model endpoint, without dropping to docker exec and kubectl patch.
Host alias commands use the legacy Kubernetes gateway Sandbox resource path.
In that older topology, the openshell-cluster-nemoclaw container runs an embedded k3s cluster with a sandboxes.agents.x-k8s.io custom resource definition, and an agent-sandbox-controller reconciles each Sandbox resource into the agent pod.
They are not supported on Docker-driver or VM-driver sandboxes because those drivers do not run the gateway cluster container that owns this resource.
The command validates the hostname and IP address, rejects duplicate hostnames, and patches spec.podTemplate.spec.hostAliases on the sandbox resource.
nemo-deepagents <name> hosts-list
List host aliases configured on the sandbox resource.
nemo-deepagents <name> hosts-remove
Remove a hostname from the sandbox hostAliases list.
nemo-deepagents <name> mcp list
List MCP servers configured for a sandbox. The command reports the selected agent’s MCP support status and, for each configured server, whether the generated OpenShell provider, policy, and agent adapter are present.
nemo-deepagents <name> mcp add
Add an MCP Streamable HTTP server to a sandbox.
Pass --url for the MCP endpoint and the required single --env KEY bearer credential for the sandbox-side MCP client.
Pass a repeatable --trusted-private-host <exact-host-or-ip> option to admit an exact RFC1918, CGNAT, or IPv6 unique local destination for the current command.
The declaration must equal the normalized host from --url.
For managed MCP, use a DNS hostname for an IPv6 unique local address because NemoClaw has not qualified direct IPv6-literal MCP URLs.
You can supply exact hosts through NEMOCLAW_TRUSTED_PRIVATE_HOSTS instead, and NemoClaw combines the variable with command options.
NemoClaw records the resulting exact trust intent and address pins, so restart, rebuild, and restore do not depend on the ambient environment.
NemoClaw registers that credential in an OpenShell provider, installs a generated OpenShell protocol: mcp policy for the target endpoint, attaches the provider to the running sandbox, and writes only an openshell:resolve:env:KEY placeholder into the agent config.
Inline --env KEY=VALUE is rejected because it would expose the value in NemoClaw process arguments.
Load the variable from a secret manager or masked prompt, export it without recording the value in shell history, and pass only --env KEY.
All endpoints must use HTTPS.
The full URL and path are persisted and displayed, so URLs cannot contain userinfo, query strings, fragments, known secret-shaped path material, percent-escaped or glob-style paths, or port zero.
Server names must start with a letter and contain at most 64 letters, digits, hyphens, or underscores, and endpoint hostnames must use canonical lowercase DNS labels.
NemoClaw rejects invalid names and endpoints before it writes lifecycle state or changes OpenShell resources.
NemoClaw generates a narrow protocol: mcp policy for the destination, literal path, adapter binaries, pinned addresses, explicit MCP methods, and a 131,072-byte request-body limit.
OpenShell 0.0.106 evaluates that policy before replacing the attached provider placeholder in the allowed request header.
NemoClaw imports the endpointless nemoclaw-mcp-v1 profile and binds the dedicated provider to that endpoint with credential_binding.provider.
OpenShell withholds the credential before the binding is active and outside the bound host, port, and path.
The sandbox client connects directly through OpenShell’s existing egress path, and NemoClaw does not run a host-side MCP data-plane bridge, proxy, relay, or listener.
After the add commits, NemoClaw freshly verifies the exact generated policy, expected provider attachment, recorded provider ID, nemoclaw-mcp-v1 type, valid resource version, and exactly one credential key matching the recorded key.
If those readiness checks pass, it sends a differential pair of wire-level MCP initialize requests from inside the sandbox — one with the placeholder header and one with an unresolvable control bearer — to verify that OpenShell resolves the credential on egress; otherwise it reports an inconclusive probe skipped result and sends no request.
Neither outcome fails the committed add, and --no-probe skips this check.
For full setup details, see Add an MCP Server.
Deep Agents MCP add and restart require the managed MCP v2 capability in the sandbox image.
If mcp add or mcp restart reports an older v1 runtime, run nemo-deepagents <name> rebuild before retrying.
NemoClaw writes managed server definitions to /sandbox/.deepagents/.nemoclaw-mcp.json; user-owned .mcp.json files are not auto-loaded by the managed harness.
For a private endpoint, use its exact URL host:
nemo-deepagents <name> mcp status
Inspect MCP server state for one server or for all configured servers.
Status includes OpenShell provider presence and credential-key shape, provider attachment, generated policy content match, adapter registration, current host-variable availability, and the selected agent’s MCP support mode.
For a trusted private endpoint, status also compares current DNS answers with recorded pins without changing the policy.
Text output reports private address pins: match, drift, or unresolved.
JSON output reports the same value in trustedPrivateTarget.state and includes the recorded pins.
When a single server is named, status requests a differential wire-level credential-resolution probe.
It sends no probe traffic unless the exact generated policy matches the effective gateway policy, the expected provider attachment is confirmed, and the live provider has the recorded ID, nemoclaw-mcp-v1 type, a valid resource version, and exactly one credential key matching the recorded key; a readiness failure reports unknown with a probe skipped detail.
When ready, the same MCP initialize is sent from inside the sandbox once with the openshell:resolve:env:KEY placeholder header and once with a deliberately-unresolvable control bearer.
Classification uses the two HTTP status codes plus curl exit codes for transport, timeout, and policy-denial outcomes; response bodies are never captured or printed.
A verified verdict requires the placeholder request to be accepted (HTTP 2xx) while the control is rejected — the only outcome that proves a valid credential was on the wire.
Identical HTTP 400, 401, or 403 rejections raise a warning that names the hypotheses — the placeholder forwarded verbatim, an expired or revoked credential that resolved correctly, or (for HTTP 400) endpoint request validation — and tells you to verify the stored credential first.
For HTTP 401 or 403, a confirmed-valid credential means the host is not rewriting placeholders and agent runtimes receive the same auth failure and skip the server; HTTP 400 remains inconclusive because the endpoint may reject the probe request itself.
Every other outcome — differing rejections (an endpoint may reject two different literal bearers differently), endpoints that skip authentication, endpoint outages, policy denials, and unreachable sandboxes — reports as unknown rather than blaming the credential rewrite, and a persisted URL that fails the current authenticated-endpoint boundary is never probed.
Pass --tools with one server name to request a live tool inventory.
The shared client runs through the managed registration’s existing OpenShell credential provider and policy; OpenShell injects the credential at that boundary, and the runtime never accepts it as an argument, environment value, or authorization option.
It performs initialize, notifications/initialized, and paginated tools/list, then attempts to close the MCP session and transport.
Cleanup errors do not replace the bounded discovery result.
It retains and returns deterministic tool names only, never prints the other tool-definition fields returned by tools/list, and never calls a tool.
The operation is bounded by total and per-request timeouts plus response-byte, page, tool-count, cursor-length, and tool-name limits.
Use --tools only with a configured endpoint you trust to advertise names while authenticated.
The endpoint controls its returned names and can derive them from the request or credential it receives; NemoClaw validates and bounds the text but cannot prove that the endpoint did not encode credential-derived data in an otherwise valid name.
The toolDiscovery JSON field contains ok, count, tools, and truncated, plus a redacted detail on failure or a bounded partial result.
These names are the server’s point-in-time advertised tools, not an attestation of the exact tools visible to the model after agent filters, progressive disclosure, or session state.
An older sandbox image without the shared client reports that the sandbox must be rebuilt.
Tool discovery is opt-in and sends authenticated network traffic to the configured endpoint.
Passing --tools suppresses the named-server credential-resolution probe that otherwise runs by default.
Pass --probe --tools to request both checks explicitly.
An unsuccessful discovery does not remove the ordinary provider, policy, environment, or adapter status from the result.
nemo-deepagents <name> mcp restart
Refresh one MCP server registration, or every server on the sandbox when no server is supplied.
Restart reapplies the generated policy, reattaches the OpenShell provider when needed, and refreshes the sandbox agent adapter registration.
For a trusted-private entry, restart replays recorded address pins without resolving the endpoint again or widening the policy.
For a public entry, restart resolves the hostname again and refreshes the policy with the current validated public addresses.
If the recorded host variable is exported, restart replaces the provider credential and waits for its new opaque revision.
Otherwise, restart reuses an existing provider whose current metadata match the registry.
A missing provider requires the variable to be exported before retrying.
An existing provider with the profile-less legacy generic type must be removed and added again with its credential exported.
OpenShell 0.0.106 cannot bind that provider to an MCP endpoint, so restart and rebuild fail closed instead of activating it.
When that provider is already absent but its name still blocks sandbox exec,
restart first detaches only the dangling sandbox-spec reference, then runs the
agent capability probe before changing a live provider or policy.
Deep Agents restart refreshes the NemoClaw-managed /sandbox/.deepagents/.nemoclaw-mcp.json projection and validates the HTTPS-only server definitions before dcode sees them.
If the sandbox still uses the older v1 MCP projection, rebuild first so restart can use the v2 capability.
nemo-deepagents <name> mcp remove
Remove an MCP server from a sandbox.
For an ordinary managed entry, NemoClaw unregisters the sandbox agent adapter, removes the exact owned generated policy, detaches and deletes the recorded OpenShell provider, and clears the sandbox registry entry.
For a stored legacy entry whose credential name is no longer accepted, it first detaches the exact provider so adapter cleanup cannot start with that credential attached.
Deep Agents teardown does not require managed MCP capability v2 from the old image.
For a v1 image, NemoClaw removes the exact registry-owned entry from the legacy .mcp.json while preserving unrelated user state; a replacement image must pass the v2 capability check before post-rebuild providers or policy are restored.
The command fails closed on observed drift.
--force may remove a modified same-name agent adapter entry, but provider deletion still requires the recorded ID and credential key plus an accepted managed provider type, and policy deletion still requires exact owned content.
An exact legacy generic provider is accepted only for cleanup.
Residuals preserve registry state.
OpenShell 0.0.106 mutates providers by name, so do not concurrently replace a managed provider through another OpenShell client during this command.
When an interrupted destroy leaves a prepared-only transaction, deletion is not durably confirmed.
If the sandbox is still live, run nemo-deepagents <name> mcp remove <server> --force with the affected server name.
NemoClaw clears the prepared marker only after cleanup succeeds without residuals and no bridge entries remain.
A failed cleanup, a wrong server name, residual resources, or any remaining bridge entry preserves the marker for another retry.
A pending marker, including a transaction with both prepared and pending markers, means the registry records that OpenShell deletion was already confirmed.
mcp remove --force refuses this state.
Run nemo-deepagents <name> destroy to finish the idempotent provider and policy cleanup.
nemo-deepagents <name> skill install <path>
Deploy a skill directory to a running sandbox.
The command validates the SKILL.md frontmatter, which requires a name field.
It uploads selected non-dot regular files while preserving their subdirectory structure.
It then performs agent-specific post-install steps.
The skill directory must contain a SKILL.md file with YAML frontmatter that includes a name field.
Skill names must contain only alphanumeric characters, dots, hyphens, and underscores.
For Deep Agents, the command installs a fresh skill directly into /sandbox/.deepagents/agent/skills/<name>, the directory Deep Agents Code loads at session start.
Before upload, NemoClaw copies each selected regular file into a private host snapshot and rejects a path that changes identity during the copy.
It creates the archive from that snapshot and records each path, normalized mode, and SHA-256 digest.
It rejects symlinks and special files.
Inside the sandbox, it stages the archive and verifies that its paths, normalized modes, and SHA-256 digests match the host snapshot.
It then moves the staged directory into place only if the destination is still absent.
On success, the command prints the content digest that the sandbox confirmed: one SHA-256 digest over the recorded paths, normalized modes, and file digests.
Record that value to compare it with the digest printed by a later install of the same skill directory.
Deep Agents Code and its built-in skill creator also write to this directory.
The command therefore refuses any name whose file, directory, or symlink already exists.
Updates are not automatic.
Use nemo-deepagents <name> connect to inspect the existing directory.
Update it manually only after confirming ownership.
The legacy /sandbox/.deepagents/skills/<name> path is not written or treated as ownership proof.
The managed dcode launchers discover newly installed skills on the next session without accepting executable hook configuration.
Installation does not enable project hooks or unmanaged MCP files.
Run nemo-deepagents <name> skill install --help to print usage for this subcommand.
If you pass a plugin-shaped directory to skill install, the CLI prints a plugin-specific hint instead of treating it as a missing skill file.
Files with names starting with . (dotfiles) are skipped and listed in the output.
Files with unsafe path characters are rejected to prevent shell injection.
Symlinks and other non-regular paths are rejected rather than followed or copied.
For OpenClaw and Hermes, an existing sandbox skill is updated in place and chat history is preserved. Deep Agents supports only fresh-name installs because its active skill directory is shared with agent-authored content. Follow the agent-specific activation guidance above after installation.
nemo-deepagents <name> skill remove <skill>
Remove an installed skill from a running sandbox by skill name when the selected agent supports automatic removal. The command validates the skill name before it applies the agent-specific removal behavior below.
For Deep Agents, automatic removal is refused before any sandbox files change.
The active /sandbox/.deepagents/agent/skills/<name> directory is shared with agent-authored content, so its presence alone cannot prove NemoClaw owns it.
Use nemo-deepagents <name> connect to inspect the existing directory.
Remove it manually only after confirming ownership.
Use the skill name from the SKILL.md frontmatter, not the local directory name.
Skill names must contain only alphanumeric characters, dots, hyphens, and underscores, and cannot be . or ...
nemo-deepagents <name> download <sandbox-path> [host-dest]
Host-side wrapper around openshell sandbox download that checks the live sandbox.
The command confirms before and after transfer that the source remains a file or directory.
Symbolic links, source-type changes, and other special source types are refused.
If the command cannot confirm the source type, it exits without publishing.
The command downloads to a fresh private temporary directory on the host, verifies that OpenShell wrote an artifact, publishes the artifact to your destination, and removes the temporary directory.
An existing destination directory is resolved to its canonical path before publication.
The command refuses an existing file destination that is a symbolic link and a new destination below a symbolic-link parent.
Regular files are published through a private temporary entry and atomically replace an existing regular file.
Relative host destinations resolve against the caller’s working directory.
Absolute host destinations do not use caller-working-directory resolution.
With no host-dest the destination defaults to the current directory.
nemo-deepagents <name> upload <host-path> [sandbox-dest]
Host-side wrapper around openshell sandbox upload, symmetric to the download wrapper.
With no sandbox-dest the destination defaults to /sandbox/ inside the sandbox.
nemo-deepagents <name> rebuild
Upgrade a sandbox to the current agent version while preserving workspace state.
The command backs up workspace state, destroys the old sandbox (including its host-side Docker image), recreates it with the current image via onboard --resume, and restores workspace state into the new sandbox.
Credentials are stripped from backups before storage.
Policy presets applied to the old sandbox are reapplied to the new one so your egress rules survive the rebuild.
Before creating the replacement sandbox, NemoClaw prints the finalized create-time policy scope whenever presets are included.
The replacement uses the recorded compatible-endpoint reasoning mode, reasoning effort, and web search selection instead of ambient shell values.
When same-gateway legacy sandbox records use the selected supported provider but omit its credential environment-variable name, rebuild fills only those missing names from the provider’s canonical configuration.
The target update and peer metadata migration use one registry update.
Conflicting credential environment-variable names, custom endpoints, or API families still stop the rebuild.
Incomplete routes and invalid gateway bindings also stop the rebuild.
NemoClaw checks the shared route again immediately before deleting the original sandbox.
Rebuild preserves the recorded sandbox GPU enablement mode and, for an explicitly enabled sandbox, its recorded device selector.
It re-resolves the Docker-driver GPU route from the current host and current NEMOCLAW_DOCKER_GPU_PATCH value, so native-only, explicitly authorized native-with-fallback, and compatibility-only routing may differ from the original onboarding run.
A rebuild preserves the recorded tool-disclosure mode unless --tool-disclosure explicitly changes it; it ignores an ambient NEMOCLAW_TOOL_DISCLOSURE value while recreating the sandbox.
A rebuild preserves the recorded Deep Agents Code observability choice and matching local OTLP policy state unless --observability or --no-observability explicitly changes them.
A rebuild preserves the recorded Deep Agents Code auto-approval capability unless --dcode-auto-approval explicitly changes it.
A sandbox onboarded with an explicit GPU opt-out (stored as sandboxGpuMode: "0", plus legacy registry entries that only record gpuEnabled: false) is recreated with the same opt-out, so the inner onboard --resume skips the Docker CDI GPU preflight on hosts without an NVIDIA GPU.
Auto-mode sandboxes remain auto.
If another terminal has an active SSH session to the sandbox, rebuild prints an active-session warning and requires confirmation before destroying the sandbox.
Pass --yes, -y, or --force to skip the prompt in scripted workflows.
The sandbox normally must be reachable for the backup step to succeed.
If an archive command preserves at least one state directory, rebuild keeps the captured backup entries and reports the manifest-defined paths that could not be archived.
If a manifest-declared state file fails, rebuild exits before destroying the original sandbox even when it preserved state directories, unless you explicitly pass --force.
If every state directory fails, rebuild exits before destroying the original sandbox even when it captured loose files, unless you explicitly pass --force.
With --force, NemoClaw preserves any captured loose files in the partial manifest and restores them after recreation.
If the backup produced nothing usable, it continues from recorded registry metadata without restoring prior sandbox state.
Use this recovery path only when losing the state that could not be backed up is acceptable.
For a sandbox with managed MCP servers, --force probes sandbox execution before MCP teardown.
If that no-op cannot run, NemoClaw requires complete bridge entries and exact live policy and provider identities, without trying an in-sandbox adapter scrub or changing MCP ownership state.
Each bridge must record the adapter for the sandbox’s recorded agent.
The registered policy must match the policy NemoClaw generates for that adapter, server name, endpoint URL, and resolved addresses.
It rechecks the registry, recorded gateway, resolved targets, live generated policies, and provider identities immediately before deletion; incomplete adds, drift, or ambiguous ownership stop before deletion.
NemoClaw sends the delete request and every deletion-confirmation lookup to the sandbox’s exact recorded gateway.
Across every rebuild path, NemoClaw does not attempt to stop local NIM until sandbox deletion is positively confirmed, then attempts NIM cleanup on a best-effort basis.
When openshell sandbox delete exits nonzero, an exact recorded-gateway lookup distinguishes explicit absence from a confirmed Ready or Running sandbox.
Any other phase or probe failure is ambiguous.
Explicit absence continues the rebuild.
Confirmed intact state triggers an attempt to restore prepared MCP state and any shields lockdown that rebuild temporarily opened.
NemoClaw reports any MCP or shields restoration failure and does not present the operation as a successful rollback.
Ambiguous state preserves MCP ownership and recovery metadata without attempting to stop NIM or claiming the original sandbox remains intact, and the rebuild process skips its immediate shields relock.
Failures after a successful exec probe do not switch to the host-side path.
Before backup or deletion, rebuild also refuses an incomplete MCP destroy transaction.
It also refuses a pending baseline exclusion transaction before opening a shields-down window, starting backup, or deleting the sandbox, and prints the exact policy exclude or policy restore command to rerun.
For a prepared-only transaction, the redacted diagnostic points to nemo-deepagents <name> mcp remove <server> --force when the sandbox is still live.
For a pending or both-marker transaction, it points to nemo-deepagents <name> destroy because the registry records that OpenShell deletion was already confirmed.
Before backup or deletion, rebuild checks the staged messaging configuration against other sandboxes in the selected OpenShell gateway’s sandbox registry.
A rebuild cannot detect messaging conflicts in an independent OpenShell gateway’s registry.
A conflict aborts with the original sandbox registered and intact so you can resolve the conflict before retrying.
After OpenShell accepts the sandbox deletion, rebuild waits until OpenShell explicitly reports that the old sandbox is absent.
Only then can NemoClaw perform any required local registry removal and begin creating the replacement.
If OpenShell does not confirm absence within the bounded wait, including when gateway transport errors block the probes, rebuild exits nonzero before registry removal or replacement creation and preserves both the local registry entry and the state backup.
Restore OpenShell connectivity and confirm the sandbox’s live state before you retry, and keep the printed backup path for recovery.
Before deletion, rebuild records a replacement journal that binds the operation to the recorded gateway, source identity, and target settings.
Rerunning the same rebuild continues from the recorded boundary or accepts the proven replacement instead of deleting it again.
A mount-free journal written before host-mount identity binding remains resumable.
An older journal that used host mounts fails closed because it cannot prove the original host source identity, even when the visible mount settings are unchanged.
Preserve the sandbox, onboarding session, printed backup, exact error, and Journaled replacement diagnostic, then follow the legacy journal guidance in Continue an Interrupted Replacement.
Use --verbose to print the replacement identifier, gateway, and journal phase.
Refer to Continue an Interrupted Replacement for the recovery procedure and fail-closed conditions.
When rebuild starts with shields up, NemoClaw opens a 30-minute shields-down window for backup and recreation.
A detached auto-lock timer remains active until NemoClaw commits a successful shields-up state, so it can attempt to restore lockdown if the host rebuild process exits unexpectedly.
After restore, the command restores Deep Agents manifest-defined state, regenerates /sandbox/.deepagents/config.toml, and recreates the managed MCP projection from the host registry.
Before changing the sandbox, rebuild verifies that the recorded inference.local route is still reachable and that the target provider, model, reasoning settings, web search selection, base image, and policy inputs match the recorded context.
If those checks fail after backup, NemoClaw restores the previous MCP state and keeps the existing sandbox intact.
Use rebuild after a failed Deep Agents version check, after enabling Tavily Search, or after upgrading from an older managed MCP runtime.
nemo-deepagents update
Check for a NemoClaw CLI update and, when requested, run the maintained installer flow. This command is a discoverable CLI wrapper around the supported installer path:
nemo-deepagents update updates the host-side NemoClaw installation.
The maintained installer flow follows the admin-promoted lkg release tag by default, so it may trail the newest semver or latest tag while validation completes.
Because of that, an install can be newer than the maintained tag.
Without --allow-downgrade, --fresh runs only when the maintained build is the same version or newer than the installed version.
When the maintained tag resolves, the command passes that repository revision to the installer, so a later tag change cannot select a different build for that update.
It reports the reason and exits non-zero in these cases:
- The installed version is newer than the maintained tag.
- The versions cannot be ordered.
- The maintained tag does not resolve to a version.
NemoClaw cannot order a git describe version against a different prerelease on the same release line.
Rerun with --allow-downgrade to reinstall regardless; --yes waives the confirmation prompt only and never accepts a downgrade on its own.
It does not replace nemo-deepagents upgrade-sandboxes; use that command to inspect or rebuild existing sandboxes after the CLI has been updated.
When the command is running from a source checkout, it reports that state and does not replace the checkout with a global package install.
nemo-deepagents upgrade-sandboxes
Rebuild sandboxes whose base image is older than the one currently pinned by NemoClaw.
NemoClaw resolves the digest of ghcr.io/nvidia/nemoclaw/sandbox-base:latest from the registry, then compares it against the digest each sandbox was created with.
Sandboxes that match the current digest are left alone.
NemoClaw also checks the build fingerprint recorded on each managed sandbox image.
A sandbox needs upgrade when its agent version is stale, when its recorded NemoClaw image fingerprint differs from the running CLI, or both.
When the target version is older than the recorded one (for example after reinstalling with an older NEMOCLAW_INSTALL_TAG), the stale listing marks the change with a (downgrade) suffix instead of framing it as a routine upgrade.
Custom Dockerfile sandboxes are not classified by image drift because rebuilding them onto the default image would drop the custom image.
Legacy sandboxes without a recorded fingerprint opt into this check after their next rebuild.
A recorded sandbox that is not observed in any phase on its own recorded gateway is reported as not found there, with remediation guidance — this typically means its gateway registration or Docker image was removed (for example by nemo-deepagents uninstall, which preserves sandboxes.json but removes both).
Before it inspects a gateway or starts a rebuild, the command validates every registered sandbox name against the NemoClaw sandbox name format.
Route-only reservations are not sandboxes and are excluded from this validation.
If the command finds incompatible names, it lists each name before any gateway inspection or rebuild.
With --check, the command then returns without changing state.
In a mutating mode, it exits with a nonzero status.
NemoClaw does not truncate or rename a registered sandbox identity.
Follow Update Sandboxes to transfer state to a compatible replacement before you rerun the command.
Each rebuild reuses the same workspace backup-and-restore flow as nemo-deepagents <name> rebuild, so workspace files survive the upgrade.
If the registry is unreachable (offline or firewalled hosts), NemoClaw falls back to the unpinned :latest tag and reports that the digest could not be resolved instead of failing.
During installer recovery, a registered sandbox that is not Ready can also be rebuilt from its validated latest backup.
That recovery requires a NemoClaw-managed image fingerprint or the installer’s explicit confirmation for a listed pre-fingerprint OpenClaw or Hermes entry.
The legacy confirmation never overrides recorded custom-image evidence.
A custom OpenClaw sandbox is recoverable only when the selected backup independently carries complete authoritative image-plugin provenance.
nemo-deepagents backup-all
Back up registered sandboxes that are running or have an eligible stopped Docker-driver container to ~/.nemoclaw/rebuild-backups/.
A registered docker-driver sandbox whose container is stopped is started for the duration of the backup and returned to its stopped state afterward.
If the container cannot be returned to the stopped state, the command fails and reports that the container was left running.
Sandboxes that are not running and cannot be started this way are skipped with remediation guidance.
For each eligible sandbox, backup-all holds one lifecycle transaction through the complete backup.
Within that transaction, it starts a stopped container when required, opens a 30-minute shields-down window when the sandbox starts with Shields up, copies sandbox state, restores the previous Shields state, and returns any container it started to the stopped state.
If the timer expires during the transaction, the deadline gate blocks new mutations and waits for the exact backup owner to finish without signaling it.
An initial lock or unlock failure marks that sandbox as failed, and backup-all continues with the next sandbox.
A failure to restore the previous Shields state stops backup-all before it processes another sandbox.
Before an OpenShell upgrade, the installer prepares the current release CLI and uses it to run backup-all in strict mode.
Strict mode requires every registered sandbox to produce a fresh backup and aborts before gateway changes if any sandbox is skipped or fails.
When strict mode reports a skipped sandbox, start that sandbox or its container and rerun the installer or nemo-deepagents backup-all.
A running sandbox whose in-sandbox SSH endpoint does not answer fails its backup and aborts the run.
For a standalone nemo-deepagents backup-all run, set NEMOCLAW_SKIP_UNREACHABLE_SANDBOX_BACKUP=1 exactly to skip such sandboxes instead of failing.
Other values such as true, yes, or 0 are not accepted.
This variable does not weaken the installer’s strict pre-upgrade requirement.
A skipped sandbox’s uncommitted state is not included in its last successful backup.
nemo-deepagents <name> snapshot create
Create a timestamped snapshot of sandbox state.
Snapshots are stored in ~/.nemoclaw/rebuild-backups/<name>/.
The command requires shields to be down and keeps the shields check and backup under one per-sandbox transition.
If the timer expires during a long-running backup, the deadline gate blocks new mutations and waits for the exact backup owner to finish.
Auto-restore does not signal the backup process.
If ownership becomes ambiguous, NemoClaw attempts to record durable containment and reports exact-generation recovery guidance.
If the containment commit fails, NemoClaw retains any exact lifecycle and deadline gates it already owns.
A state-directory failure that prevented gate publication also prevents normal mutation-lock acquisition.
Correct the reported state-directory write failure, then run nemo-deepagents <name> shields status to resume recovery or receive exact-generation recovery guidance.
When the sandbox has active baseline exclusions, successful output lists their keys and repeats that excluded egress leaves dependent agent features unsupported for that sandbox.
Names must be 1 to 63 characters from [A-Za-z0-9._-], start with an alphanumeric character, and cannot look like a version selector (v1, v2, …). Duplicate names per sandbox are rejected; pick a different name or delete the existing snapshot first.
When a directory or state file cannot be captured, the command reports the failed items, removes the incomplete snapshot, and exits nonzero.
A removed snapshot does not appear in snapshot list and cannot be restored, so a later restore cannot select a capture that never completed.
When removal fails, the command reports the listed snapshot path and exits nonzero.
Remove that directory manually before you run snapshot restore because the incomplete capture remains selectable.
nemo-deepagents <name> snapshot list
List available snapshots for a sandbox as a table of version, name, timestamp, and path.
Versions (v1, v2, …) are computed on read from timestamp-ascending order, so v1 is the oldest snapshot and vN is the newest. Snapshots created before this feature landed are numbered retroactively.
nemo-deepagents <name> snapshot restore [selector] [--to <dst>] [--force] [--yes|-y]
Restore sandbox state from a snapshot.
For an in-place restore, the sandbox must be running.
If no selector is provided, the latest snapshot is used.
Restore removes files added after the snapshot only from state directories selected for cleanup.
It preserves directories that exist only in the target manifest or whose backup failed.
The state replacement, mutable-config permission repair, and policy reconciliation run under the same per-sandbox transition.
If the timer expires during that work, the deadline gate blocks new mutations and waits for the exact restore owner to finish.
Auto-restore does not signal the restore process.
If ownership becomes ambiguous, NemoClaw attempts to record durable containment and reports exact-generation recovery guidance.
If the containment commit fails, NemoClaw retains any exact lifecycle and deadline gates it already owns.
A state-directory failure that prevented gate publication also prevents normal mutation-lock acquisition.
Correct the reported state-directory write failure, then run nemo-deepagents <name> shields status to resume recovery or receive exact-generation recovery guidance.
Post-restore policy reconciliation is best-effort. NemoClaw warns and continues the remaining restore steps in these cases:
- NemoClaw cannot verify whether a custom policy owns the live
observability-otlp-localpolicy entry. - The built-in
observability-otlp-localpolicy preset has drifted or cannot be inspected. - NemoClaw cannot add or remove a recorded policy preset.
The live network policy can then retain unwanted egress or omit expected egress until you repair the named preset.
After a warning, run nemo-deepagents <name> policy list.
Confirm that the named preset is recorded in the sandbox registry and active on the gateway, or absent from both.
The selector accepts any of:
- A version (
v1,v2, …,vN) fromsnapshot list. - An exact name passed to
snapshot create --name. - An exact timestamp.
Pass --to <dst> to restore the snapshot into a different sandbox instead of the source.
When dst does not exist, it is auto-created from the source image.
For a Docker- or VM-driver source, the source can be stopped when its registry entry records both the sandbox image and a complete inference route.
For a Kubernetes-driver source, the pod image must remain resolvable through its gateway.
No re-onboarding is needed when those prerequisites are present.
For a new destination, NemoClaw requires its owning gateway to report Ready state and a valid live identity.
It revalidates that identity immediately before registration.
The destination receives a new lifecycle generation and does not inherit the source sandbox’s generation.
If the destination is not Ready with the same valid identity, the command exits nonzero before registration or state restore.
The created destination remains unregistered, so --force cannot select it for deletion.
Run the exact owner-scoped deletion command printed by the failure:
After OpenShell deletes the destination, rerun the original snapshot restore --to command.
A cross-sandbox restore refuses to clone a source or replace an existing destination whose baseline exclusion transaction needs repair, before it creates or deletes anything.
When dst already exists, snapshot restore --to <dst> refuses by default to avoid silently mutating the destination’s filesystem.
To overwrite an existing destination, pass --force: the command deletes dst, then recreates it from the source’s image and restores the snapshot into the fresh copy.
If the existing destination has an active shields timer, the force path restores and verifies lockdown, revokes the timer, and then deletes the destination.
It clears the remaining local shields state only after deletion succeeds.
The --force path prompts interactively to confirm the destination name before deleting.
Pass --yes (or set NEMOCLAW_NON_INTERACTIVE=1) to skip the prompt.
The snapshot selector, source image, and durable inference route are validated before any deletion. If any prerequisite is invalid, restore stops before it deletes dst.
When --to names an existing sandbox, restore refuses to overwrite it unless you pass --force.
With --force, NemoClaw confirms the destructive restore unless you also pass --yes or run with NEMOCLAW_NON_INTERACTIVE=1.
Use this path only when the destination sandbox can be replaced by the selected snapshot.
nemo-deepagents <name> share mount
Mount the sandbox filesystem on the host machine via SSHFS for bidirectional file sharing. Files edited on the host appear instantly inside the sandbox, and vice versa.
Expected output:
Prerequisites:
sshfsmust be installed on the host (sudo apt-get install sshfson Linux,brew install macfuse && brew install sshfson macOS).- The sandbox must be running.
- The remote sandbox path must exist. NemoClaw verifies it against the target sandbox before invoking
sshfsand prints aconnect, thenls <path>check when the probe fails. - Sandboxes created before the
openssh-sftp-serverbase image update must be rebuilt withnemo-deepagents <name> rebuild. - The local mount path must be on a writable filesystem; FUSE creates the mount on the host side.
If the default
~/.nemoclaw/mounts/<name>lives on a read-only filesystem, pass an explicit writable path as the second positional argument.
nemo-deepagents <name> share unmount
Unmount a previously mounted sandbox filesystem.
nemo-deepagents <name> share status
Check whether the sandbox filesystem is currently mounted.
Expected output:
openshell term
Open the OpenShell TUI to monitor sandbox activity and approve network egress requests. Run this on the host where the sandbox is running.
For a remote server, connect through SSH and run openshell term on that server.
nemo-deepagents status
Show the global sandbox list and the status of host auxiliary services (for example cloudflared).
This command is host-wide. It summarizes registered sandboxes, the default sandbox’s live inference route, gateway health, and host services.
Use nemo-deepagents <name> status when you need one sandbox’s live health and recovery guidance.
Pass --json for machine-readable output with registered sandboxes, service state, inference routes, and health details.
Each JSON sandbox row reports agent as a string, never null.
The row reports openclaw when the registry records no agent for the sandbox.
This command reads the registry without gateway recovery, so it never reports unknown.
For each listed sandbox, the text output includes the configured inference provider and model plus the number of active SSH sessions when the session probe is available.
Host-service PID lookup honors NEMOCLAW_SANDBOX_NAME, then NEMOCLAW_SANDBOX, then SANDBOX_NAME, then the registry default.
When at least one sandbox is registered and the named NemoClaw gateway is unreachable, unhealthy, or attached to a different sandbox, the command prints a gateway: down [state] (reason) line between the sandbox list and the host-service list.
The command classifies the failing layer when possible: the named gateway port is not accepting connections, the named gateway is running but not Connected, the active OpenShell gateway points at a different name, or the named gateway is not configured at all.
It then prints the gateway recovery guidance for your host.
That guidance names nemo-deepagents onboard when NemoClaw starts the gateway process.
When another deployment owns that process, the guidance directs you to start it with the owning deployment and run openshell gateway select <gateway>.
It exits with code 1 so shell scripts and CI can detect the degraded state from $?.
For --json, the structured output includes gatewayHealth, and the exit code is set after the report is generated.
A clean machine with no registered sandboxes keeps the legacy 0 exit because no gateway is expected to be configured yet.
If cloudflared is installed but not running, the host-service section reports whether the PID file is missing, invalid, or points at a dead process, then suggests nemo-deepagents tunnel start as the recovery command.
nemo-deepagents inference get
Show the active live inference provider and model from the NemoClaw-managed OpenShell gateway.
Use this command when you want the direct runtime route without the rest of the sandbox status output.
It is also available in sandbox-first form as nemo-deepagents <name> inference get.
The sandbox-first grammar nemo-deepagents <name> inference get is also accepted and reads the same gateway-wide route, so it stays symmetric with nemo-deepagents <name> inference set.
nemo-deepagents inference set
For Deep Agents sandboxes, run nemo-deepagents onboard --fresh --name <sandbox-name> --recreate-sandbox when you need to change the provider or model.
The managed dcode configuration is written under /sandbox/.deepagents during onboarding, so the recreate path keeps the OpenShell route and the sandbox config aligned.
Use nemo-deepagents inference get and nemo-deepagents <name> status to inspect the current route.
nemo-deepagents setup
The nemo-deepagents setup command is deprecated.
Use nemo-deepagents onboard instead.
This command remains as a compatibility alias to nemo-deepagents onboard and accepts the same flags: --profile <name>, --non-interactive, --resume, --fresh, --recreate-sandbox, --gpu / --no-gpu, --from, --name, --host-mount, --sandbox-gpu / --no-sandbox-gpu, --sandbox-gpu-device, --agent, --agents <agents.yaml>, --tool-disclosure <progressive|direct>, --observability / --no-observability, --control-ui-port, --yes / -y, --no-ollama-autostart, --yes-i-accept-third-party-software.
nemo-deepagents setup-spark
The nemo-deepagents setup-spark command is deprecated.
Use the standard installer and run nemo-deepagents onboard instead, because current OpenShell releases handle the older DGX Spark cgroup behavior.
This command remains as a compatibility alias to nemo-deepagents onboard and accepts the same flags: --profile <name>, --non-interactive, --resume, --fresh, --recreate-sandbox, --gpu / --no-gpu, --from, --name, --host-mount, --sandbox-gpu / --no-sandbox-gpu, --sandbox-gpu-device, --agent, --agents <agents.yaml>, --tool-disclosure <progressive|direct>, --observability / --no-observability, --control-ui-port, --yes / -y, --no-ollama-autostart, --yes-i-accept-third-party-software.
nemo-deepagents debug
Collect diagnostics for bug reports.
Gathers system info, Docker state, gateway logs, and sandbox status into a summary or tarball.
Use --sandbox <name> to target a specific sandbox, --quick for a smaller snapshot, or --output <path> to save a tarball that you can attach to an issue.
If --output is set and the tarball cannot be written (for example, the destination directory is missing or read-only), the command exits non-zero so scripts can detect the failure.
The tarball is written to a temporary sibling and renamed on success, so a pre-existing file at --output is preserved when tar fails.
When --sandbox is supplied explicitly through the flag or one of NEMOCLAW_SANDBOX_NAME, NEMOCLAW_SANDBOX, or SANDBOX_NAME, the name must match a registered sandbox.
The flag wins, then the env vars in that order.
If openshell sandbox list succeeds, the sandbox must also appear in the live gateway.
An unknown or stale name exits non-zero with an actionable error that names the sandbox and reports the source env var when applicable, and no tarball is written.
Without an explicit name, nemo-deepagents debug falls back to the registry’s default sandbox and warns if that default is stale.
nemo-deepagents credentials list
List the provider credentials registered with the OpenShell gateway. Values are not printed.
nemo-deepagents credentials add <PROVIDER>
Register a provider credential with the OpenShell gateway by name and type.
Each --credential takes the env variable name whose value the gateway should read; export the value first so it is not placed in argv.
Pass either repeatable --credential <ENV_NAME> or --from-existing, but do not combine them.
--from-existing is available only when no managed MCP server reserves credential keys.
The command fails before gateway work when a reservation exists because --from-existing does not expose credential keys before provider creation.
Rerun with explicit --credential <ENV_NAME> input, or remove every managed MCP server that reserves credential keys before retrying.
After the gateway accepts the provider, rebuild the target sandbox so the new provider is attached.
Registered providers attach to every sandbox you build or rebuild after the call (the gateway is one process serving all sandboxes).
If you want a provider available to only some sandboxes, scope it with nemo-deepagents credentials reset <PROVIDER> once those sandboxes finish using it.
nemo-deepagents credentials reset <PROVIDER>
Remove a provider credential from the OpenShell gateway by provider name.
After removal, re-running nemo-deepagents onboard re-prompts for that provider’s credential.
Run nemo-deepagents credentials list first if you are not sure of the provider name.
nemo-deepagents gc
Remove orphaned sandbox Docker images from the host.
Sandbox creation can build images in the gateway-managed openshell/sandbox-from repository or the locally prebuilt nemoclaw-sandbox-local repository.
The destroy and rebuild commands clean up the image automatically, but images from older NemoClaw versions or interrupted operations may remain.
This command lists images from both repositories, cross-references the sandbox registry, and removes any that are no longer associated with a registered sandbox.
nemo-deepagents uninstall
Run uninstall.sh to uninstall NemoClaw. Unless this section explicitly describes portable cleanup, its resource-removal statements apply outside portable cleanup.
The CLI runs the local uninstall.sh shipped with the installed npm package.
If that local script is missing, the CLI does not auto-fetch a remote copy.
It prints the versioned URL of the matching uninstall.sh so you can download, review, and run it manually.
When the gateway is externally supervised, uninstall preserves its process, Docker resources, and OpenShell binaries.
It still deletes the selected sandboxes and attempts to remove the modern local gateway registration.
When uninstall confirms that no sibling gateways remain, it also deletes NemoClaw provider registrations.
For a managed dual-Station vLLM runtime, full uninstall revalidates the exact recorded pair and removes both managed containers before starting the remaining uninstall steps.
If that cleanup fails, uninstall exits nonzero, preserves its owner-only cleanup receipt, and tells you to resolve the reported peer error before retrying.
Pair cleanup can partially complete before an error; verify both Stations before the retry.
For an authenticated host-local vLLM runtime, full uninstall verifies the exact named container, NemoClaw ownership label, persisted API key, and authentication fingerprint before removing the container by its inspected ID.
When that ownership state is missing, full uninstall removes the reserved nemoclaw-vllm container only when Docker reports its NemoClaw managed label and a valid container ID.
An unlabeled container or malformed inspection remains in place and stops the remaining uninstall steps.
For managed llama.cpp, full uninstall verifies the exact named container and network ownership before removing both resources by their inspected IDs.
These host-local checks run before NemoClaw deletes their state.
If Docker is unavailable or a resource does not match its persisted ownership state, uninstall exits nonzero before the remaining uninstall steps and preserves that state for recovery.
Host-local cleanup can partially complete before an error.
Restore Docker access or resolve the named ownership conflict, inspect the remaining container and network, and retry uninstall.
Managed llama.cpp and vLLM model files remain in the shared Hugging Face cache by default.
Outside portable cleanup, --delete-models deletes every model in the local Ollama inventory and all non-credential data in the current user’s shared ~/.cache/huggingface/ cache.
This opt-in can delete cached files that other applications installed or use.
It preserves the Hugging Face token and stored_tokens authentication files.
NemoClaw stops and verifies its managed local and distributed model runtimes before it deletes non-credential data from the local Hugging Face cache.
It does not scan arbitrary directories or delete model caches on remote peers.
When sibling gateway environments remain, uninstall preserves both model stores even if you pass --delete-models.
An Ollama inventory error, model deletion error, unsafe cache path, or cache-data deletion error makes uninstall exit nonzero.
Cleanup can partially complete before an error, so resolve the reported error and rerun uninstall.
It does not use the legacy gateway destroy command for that gateway.
Outside portable cleanup, uninstall also stops any orphaned openshell host processes left behind by previous onboard or destroy cycles, including openshell sandbox create, openshell ssh-proxy, and SSH sessions spawned by OpenShell.
Earlier releases only stopped openshell forward processes, so those orphans accumulated across runs.
Outside portable cleanup, uninstall also stops matching Ollama auth proxy processes before deleting ~/.nemoclaw state so stale proxy listeners do not block a later reinstall.
When sibling gateways remain, uninstall leaves the shared proxy running for them.
For Hermes setups, uninstall inspects the selected gateway’s managed port-forward watcher state, stops each verified watcher process and its sandbox-scoped forward, and leaves sibling gateway state untouched. If any watcher or forward cleanup cannot be confirmed, uninstall exits nonzero and preserves the selected gateway’s watcher state so you can retry cleanup.
Outside portable cleanup, Linux uninstall removes ~/.local/state/nemoclaw unless you pass --keep-openshell, the gateway is externally supervised, or another gateway-port environment remains on the host.
That directory contains NemoClaw-owned Docker-driver gateway configuration and SQLite data, audit logs, VM-driver state, and standalone-fallback gateway PID files.
Uninstall preserves it when the managed or externally supervised gateway process remains because that process depends on the state.
When another gateway-port environment remains, uninstall removes only the selected gateway port’s subdirectory of that directory and keeps the other ports’ subdirectories.
Run nemo-deepagents uninstall --all-gateway-ports to remove every gateway port on the host.
Keep a declared external gateway state directory outside that NemoClaw-owned path.
Uninstall does not otherwise target the declared external directory.
NEMOCLAW_GATEWAY_PORT selects the gateway instance and state root to uninstall.
Port 8080 selects nemo-deepagents and the shared ~/.nemoclaw/ root; a non-default port selects nemoclaw-<port> and ~/.nemoclaw/gateways/<port>/.
For example, NEMOCLAW_GATEWAY_PORT=9123 nemo-deepagents uninstall selects nemoclaw-9123.
The compatibility --gateway flag cannot select another instance: when present, it must match the name derived from NEMOCLAW_GATEWAY_PORT, or uninstall exits before cleanup.
Default-port uninstall removes NemoClaw-managed entries in openshell/gateway.env.
For a NemoClaw-managed authority, it also removes only NemoClaw’s marked Linux gateway unit.
It preserves upstream Linux package units, the macOS Homebrew service, and unrelated environment entries.
Gateway-scoped cleanup removes that gateway’s OpenShell resources first, then the marked Linux unit.
The OpenShell gateway service therefore keeps running while uninstall deletes the selected gateway’s sandboxes.
If OpenShell resource cleanup fails, uninstall exits nonzero and preserves the marked Linux unit and gateway process.
If marked Linux unit cleanup fails, uninstall exits nonzero before it scans for or stops a remaining gateway process or continues with later Docker and gateway-state cleanup.
OpenShell resource and Linux unit cleanup can partially complete before either failure.
After selected sandbox cleanup succeeds, uninstall removes those entries from sandboxes.json before gateway registration and Linux unit cleanup.
If a later step fails, the retry skips the completed sandbox deletions and resumes the remaining cleanup.
Resolve the reported error.
Inspect the remaining gateways with openshell gateway list.
Rerun NEMOCLAW_GATEWAY_PORT=<port> nemo-deepagents uninstall with the gateway port from the failed uninstall.
For an externally supervised authority, uninstall preserves the selected local gateway state in both full and gateway-scoped cleanup.
It also preserves the gateway process, supervisor resources, marked Linux unit, Docker resources, OpenShell binaries, and the declared external state directory.
A custom-port uninstall does not stop or remove the default gateway service or its environment file.
Uninstall does not stop an openshell-gateway process that another non-root user owns and that this installation did not record.
It names the owner and process ID, leaves that process running, and continues with the remaining cleanup.
If no other cleanup fails, uninstall exits with status 0 even though that process can keep its port in use.
Uninstall still tries to stop a root-owned process and the gateway process that this installation recorded.
If either stop fails, uninstall reports the process without printing a reusable privileged kill command.
Do not signal a PID from saved output.
Immediately before a privileged stop, verify that the live process owner and openshell-gateway command line match the exact gateway name and port.
Also prove that the PID file, runtime marker, and loaded sandbox namespace still match the selected state directory.
Rerun uninstall after the process stops.
A gateway-scoped uninstall and every --all-gateway-ports pass exit nonzero after that failure.
A single full uninstall reports the process and continues.
Before each sandbox deletion during scoped Docker cleanup, NemoClaw proves the selected configuration and running gateway identity again and passes the selected gateway name to OpenShell.
The configuration and running process must use the state-root-specific OpenShell sandbox namespace that NemoClaw generated.
For a NemoClaw-managed gateway, the live proof also binds the process owner, PID file, runtime marker, and command line to the exact gateway name and port.
For an externally supervised gateway, NemoClaw proves the configured state.
It binds the supervisor’s current main process to its owner, loaded sandbox namespace, declared executable, selected gateway name, and selected port.
When NemoClaw can prove an owner-private, generated configuration and complete JWT bundle that predate state-root scoping, restart keeps the legacy gateway ID, JWT bundle, and Docker driver’s default namespace.
That compatibility keeps the gateway able to find existing containers and keeps their non-expiring sandbox JWT issuer valid.
NemoClaw regenerates the other gateway settings from the current runtime configuration.
For a proven legacy Podman gateway, NemoClaw preserves the gateway ID that existing sandbox JWTs use; the supported Podman schema has no sandbox_namespace setting to preserve.
If the existing identity is ambiguous or unsafe, or durable gateway state remains without its configuration, restart fails closed without rewriting the configuration or JWT bundle.
Fresh state roots and already scoped configurations continue to use the state-root-specific identity.
The legacy default namespace is not isolated across gateways, so it cannot satisfy the scoped-uninstall proof while sibling gateways remain.
Scoped uninstall stops before it deletes a sandbox, registry row, or gateway registration and preserves the selected gateway’s runtime evidence and local state.
Because the supported OpenShell Podman schema does not expose sandbox_namespace, scoped Podman uninstall fails closed before signaling and preserves the gateway runtime evidence and local state.
Full single-gateway Podman uninstall for the portable experimental profile uses exact receipt-owned durable retirement and preserves shared OpenShell providers, the gateway service and process, binaries, configuration, model stores, and Podman images under the portable authority, retry, and preservation contract.
For Docker, if any proof is absent, uninstall exits nonzero before it signals the host gateway.
NemoClaw preserves the gateway runtime evidence and local state.
Keep that state intact.
For an already scoped gateway with stale runtime evidence, restore it through the supported install or onboarding recovery flow, verify the generated identity, and retry.
A proven legacy gateway is not silently converted by onboarding.
To retire one, first remove sibling gateways through their own proven scoped cleanup, verify that only the legacy gateway remains, and then use the full single-gateway uninstall path.
For an ambiguous or incomplete identity, stop the gateway and restore the exact generated openshell-gateway.toml and complete jwt/ directory from a dedicated host-level backup path, represented here as <gateway-identity-backup>.
The backup must have been captured from that gateway’s exact state directory before the failure and kept under the owning user’s exclusive access.
Keep the <gateway-identity-backup> directory and its nested jwt/ directory at mode 0700, and keep the configuration and JWT files at mode 0600.
The default gateway stores them under ~/.local/state/nemoclaw/openshell-docker-gateway/; a non-default gateway uses ~/.local/state/nemoclaw/openshell-docker-gateway-<port>/.
Restore them as the owning user.
Keep the gateway state root and its jwt/ directory at mode 0700, and do not grant group or other access to the configuration or JWT files.
NemoClaw does not reconstruct gateway identity from sandbox snapshots or backup-all; if no matching gateway-state backup exists, keep the state intact rather than attempting a scoped cleanup.
Verify every gateway with openshell gateway list.
Retain <gateway-identity-backup> only until that command reports the restored gateway and the affected existing sandboxes authenticate successfully.
Then remove that dedicated backup directory as the owning user and verify its absence by replacing the placeholder in test ! -e '<gateway-identity-backup>' with the full backup path.
If verification fails, keep the backup under the same access restrictions and stop.
Do not add sandbox_namespace manually to a live gateway configuration because the running process can still be using its previous namespace.
Uninstalling Every Gateway Port
A single uninstall is scoped to one gateway port, so the other ports on the host keep running and keep their ports bound.
When uninstall detects other gateway-port environments, it names each one, gives the NEMOCLAW_GATEWAY_PORT=<port> command that removes one of them, and points at the whole-host sweep.
A gateway environment whose port cannot be read is reported as an unidentified environment rather than omitted.
--all-gateway-ports, or NEMOCLAW_UNINSTALL_ALL_GATEWAY_PORTS=1, uninstalls all of them in one run.
The sweep enumerates the default state root and the non-default roots under ~/.nemoclaw/gateways/.
When the sweep finds more than one port, it confirms once against the resulting port list, then uninstalls each other port before the port NEMOCLAW_GATEWAY_PORT selects.
When it finds only the selected port, it uses the standard uninstall confirmation without a port list and runs that port once.
Each port runs as its own uninstall so that every port-scoped value, including the state root, registry file, gateway name, and Docker resource names, resolves from that port rather than from the calling environment.
Outside portable cleanup, the selected port runs last so its pass can remove the shared host resources once no other environment remains.
--delete-models, --destroy-user-data, and --keep-openshell apply to every port, subject to the portable preservation contract; --gateway remains a check against the selected port only.
A failure to enumerate the gateway state roots safely stops the sweep before any port uninstall begins.
The sweep cannot select an unidentified environment until its gateway port can be determined.
A port that fails to uninstall is reported, the sweep continues, and the exit code is nonzero.
That port still counts as a live sibling, so the final pass falls back to gateway-scoped cleanup and preserves the shared host resources.
Cleanup that completed before a port failure is not rolled back.
Resolve the reported error, inspect the remaining gateways with openshell gateway list, and rerun the sweep or the named per-port command.
User-data preservation under ~/.nemoclaw/
To avoid uninstall destroying host-side user data, uninstall preserves the following entries in the selected gateway’s state root by default.
The default gateway uses ~/.nemoclaw/; a non-default gateway uses ~/.nemoclaw/gateways/<port>/.
Outside portable cleanup, when uninstall confirms that no sibling gateways remain, it also removes shared host resources such as the gateway source clone, runtime state, and the Ollama auth proxy PID file.
When sibling gateways remain, it removes only the selected gateway’s resources and port-scoped state while preserving those shared host resources.
With --destroy-user-data, that scoped path removes installer-managed user-local CLI shims under ~/.local/bin/ only when sibling evidence is unidentified (for example odd ~/.nemoclaw/gateways/ entries or an unreadable gateway list). When a confirmed sibling gateway port remains, those shared shims stay with the shared npm CLI package and the other shared host resources.
If the OpenShell command is unavailable or its gateway list cannot be read, uninstall cannot confirm that the selected gateway is the last one, so it uses the same scoped path and preserves the shared resources.
When the command itself is unavailable, uninstall exits nonzero before OpenShell cleanup so you can restore the command and retry.
--yes deliberately remains non-destructive for user data.
It only acknowledges the global Proceed? confirmation prompt and still preserves the listed entries.
Removing the preserved entries always requires an explicit opt-in flag (--destroy-user-data) or the matching env var (NEMOCLAW_UNINSTALL_DESTROY_USER_DATA=1).
Existing automation using --yes therefore retains its safe behaviour and never loses host-side state by accident.
Decision matrix:
The preserved entries survive uninstall as inert files on disk.
Reinstall NemoClaw and re-onboard the sandbox before nemo-deepagents <name> snapshot restore can use them.
The preserved sandboxes.json file does not make the recorded sandboxes recoverable on its own.
Uninstall deletes the selected sandboxes and attempts to remove the local gateway registration.
Outside portable cleanup, after uninstall confirms that no sibling gateways remain, it also deletes provider registrations.
Outside portable cleanup, a NemoClaw-managed gateway also removes the Docker image.
For an externally supervised gateway, it preserves Docker resources, but the registry still cannot recover deleted sandbox and provider resources.
Uninstall warns about this at preserve time.
After reinstalling, the installer reports such records as not found on their recorded gateway instead of claiming they were recovered; run nemo-deepagents <name> destroy to clear a stranded record, then nemo-deepagents onboard to rebuild it.
Pass --destroy-user-data at uninstall time if you prefer to purge the registry along with its dependencies.
nemo-deepagents uninstall vs. the hosted uninstall.sh
Both forms execute the same uninstall.sh with the same flags, but differ in where the script comes from and how much they trust the network.
Use nemo-deepagents uninstall by default.
Use the hosted curl … | bash form only when the CLI is broken or already partially removed.
Internal Commands
NemoClaw registers a hidden internal command namespace. These commands are
compatibility entrypoints for repo-owned scripts, such as the installer, the
uninstaller, DNS setup, and developer tooling. They are not part of the
supported public CLI surface.
Each command class sets hidden = true, so the commands stay out of
nemo-deepagents --help. They remain registered and routable, which is why they are
listed here for reference. Treat their names, flags, and output as
implementation details. They exist to back install.sh, uninstall.sh, and
related automation, and they may change or be removed without notice. Most run
indirectly through those scripts rather than being typed by hand.
For contributor guidance on how these command files are structured, refer to
src/commands/internal/README.md.
These commands do not appear in the command-level parity check, which compares
nemo-deepagents --help against the public command headings in this reference; hidden
commands are excluded from both. The table above is the canonical reference for
the script-backed family.
The experimental adapter is documented separately because it has no owning script.
nemo-deepagents internal voice-gateway serve is registered for the OpenClaw-only experimental adapter described below.
Hermes and Deep Agents Code do not have an equivalent adapter.
The experimental voice gateway has no Hermes or Deep Agents Code equivalent.
Environment Variables
NemoClaw reads the following environment variables to configure service ports, onboarding behavior, and lifecycle defaults.
Set them before running nemo-deepagents onboard or any command that starts services.
All ports must be non-privileged integers between 1024 and 65535, unless a variable’s own description gives a narrower range.
CLI Logging
The centralized CLI logger writes its output to stderr and uses info verbosity by default.
These controls affect leveled logger output; they do not suppress command results or command-specific output that has not migrated to the centralized logger.
The environment precedence is NEMOCLAW_LOG_LEVEL, then NEMOCLAW_DEBUG, followed by the default info level.
The error level prints errors only, warn also prints warnings, info also prints informational messages, and debug prints all levels with timestamps.
Use these NemoClaw-specific variables instead of the generic DEBUG variable. DEBUG is not a NemoClaw logger control and can enable dependency diagnostics that include raw command arguments.
Commands whose parser owns the base logging options also accept the hidden long-form --debug and --quiet flags, even though these options do not appear in command help.
The flags are mutually exclusive.
--debug overrides the environment-derived threshold and selects debug, while --quiet caps verbosity at warn without increasing an environment-derived error threshold.
There is no global -q logging shorthand.
Passthrough commands do not consume flags intended for the downstream command as host logging options, so use the environment variables when you need unambiguous host logging around a passthrough invocation.
If a port value is not a valid integer or falls outside the allowed range, the CLI exits with an error.
NEMOCLAW_GATEWAY_PORT also cannot overlap configured service, vLLM, Ollama, Ollama proxy, Bedrock Runtime adapter, OpenRouter runtime adapter, or HTTPS Pin Runtime adapter ports, and cannot use reserved auto-allocation ranges or the default inference/proxy ports 8000, 8081, 11434, 11435, 11436, 11437, and 11438.
Port 8081 is reserved for authenticated existing-server attachment and the managed llama.cpp runtime.
It cannot be assigned to another configurable NemoClaw service port.
Each runtime adapter port must be distinct from the gateway, vLLM, Ollama, Ollama proxy, dashboard allocation range, and other runtime adapter ports.
When you run multiple NemoClaw gateways with different NEMOCLAW_GATEWAY_PORT values, NemoClaw derives a separate gateway name, state directory, and compatibility container name from the port so one gateway does not tear down another.
Only port 8080 uses a NemoClaw-managed Linux systemd user service or macOS Homebrew service.
NemoClaw-managed gateways on custom ports run as detached processes and do not change the default gateway service.
An externally supervised gateway can use any matching configured port and must be recovered through its declared supervisor.
On non-WSL hosts, NEMOCLAW_OLLAMA_PORT and NEMOCLAW_OLLAMA_PROXY_PORT must be different.
If you run Ollama on port 11435, set NEMOCLAW_OLLAMA_PROXY_PORT to another free port before onboarding.
NEMOCLAW_GATEWAY_BIND_ADDRESS accepts only 127.0.0.1 and 0.0.0.0, but NemoClaw rejects 0.0.0.0 for Docker-driver gateways while gateway JWT auth is active.
These overrides apply to onboarding, status checks, health probes, and the uninstaller. Defaults are unchanged when no variable is set.
Onboarding Configuration
The following variables let you tune onboarding without editing the Dockerfile or passing repeated flags.
Set them before running nemo-deepagents onboard.
Linux Ollama install mode details
Set NEMOCLAW_OLLAMA_INSTALL_MODE=system to run the official https://ollama.com/install.sh installer, which uses sudo, writes to /usr/local, and configures systemd.
Set NEMOCLAW_OLLAMA_INSTALL_MODE=user to extract the release tarball to ${HOME}/.local without sudo and launch the daemon manually without systemd persistence.
Leave NEMOCLAW_OLLAMA_INSTALL_MODE empty or unset to let NemoClaw auto-detect the mode.
Auto-detection selects system when the current user is root or passwordless sudo works.
Auto-detection selects user in non-interactive runs without passwordless sudo.
An interactive shell falls back to system so the official installer can prompt for the password.
NemoClaw rejects any other value.
On upgrades, NemoClaw rejects user because a user-local install cannot replace the system daemon on :11434.
On upgrades, NemoClaw also rejects system under NEMOCLAW_NON_INTERACTIVE=1 when passwordless sudo is unavailable because the installer would hang on a hidden sudo prompt.
The run exits with an actionable diagnostic instead.
Experimental NemoCUA
Keep NEMOCLAW_CUA_ENABLED=1 set whenever NemoClaw uses the experimental nemocua agent, including discovery, launch, agent commands, sandbox creation, and rebuild.
Onboarding Behavior Flags
The following flags toggle optional behaviors during onboarding.
Set them before running nemo-deepagents onboard.
Set NEMOCLAW_LANGCHAIN_DEEPAGENTS_CODE_SANDBOX_BASE_IMAGE_REF to a LangChain Deep Agents Code sandbox-base tag or digest to override base-image resolution during onboarding.
NemoClaw requires environment overrides to use the official remote repository and resolve to a repository digest, then validates the requested image against the manifest-required deepagents-code package version before using it.
NemoClaw accepts local bases only when it builds and pins them during onboarding.
Onboard Profiling Traces
Set NEMOCLAW_TRACE=1 before nemo-deepagents onboard to write an OpenTelemetry-style JSON trace for the run.
If you do not set a trace path, NemoClaw writes a timestamped file under .e2e/traces/ in the current working directory.
Use NEMOCLAW_TRACE_DIR to choose the output directory, or NEMOCLAW_TRACE_FILE to choose the output file.
Trace artifacts include onboard phase timing, sandbox and service readiness waits, policy application, inference validation probes, curl probe results, and sandbox build progress events. Secret-like metadata such as API keys, bearer tokens, cookies, and credentials is redacted before the file is written.
Deep Agents Code OTLP Traces
Pass --observability during Deep Agents onboarding to enable backend-neutral runtime traces for Deep Agents Code.
This feature is separate from NEMOCLAW_TRACE, which records NemoClaw onboarding phases, and from the OpenClaw diagnostics plugin.
The sandbox sends OTLP/HTTP protobuf requests only to http://host.openshell.internal:4318/v1/traces.
The managed exporter uses standard OTLP transport headers but does not accept operator-supplied custom or authentication headers.
A host operator must run the receiver on port 4318 and configure any Jaeger, Phoenix, LangSmith, or other backend exporter on the collector side.
Changing the host collector’s exporter does not require a sandbox rebuild or policy change.
Collector and exporter failures are non-fatal to agent work.
Native LangSmith tracing and ambient OTLP configuration remain disabled in the sandbox. The explicit opt-in can export bounded prompts, responses, tool arguments, tool results, and operational metadata, so operators must treat trace payloads as sensitive application data. The collector must enforce the operator’s filtering and redaction requirements before remote forwarding because the local policy applies to the managed Python interpreter and does not provide authenticated tenant identity. For a runnable LangSmith collector setup, refer to Set Up Deep Agents Trace Export. For the receiver trust contract, refer to Understand Deep Agents Trace Export.
Probe Timeouts
The following variables tune how long internal probes wait before giving up. Defaults are sized for typical hardware; override only if you see false-positive timeouts.
Onboard and Sandbox Readiness Timeouts
The following environment variables tune onboard-time and recovery wall-clock limits.
Set the onboarding variables before running nemo-deepagents onboard if a slow connection or large model pull risks tripping the default.
An unset, blank, invalid, or negative NEMOCLAW_GATEWAY_RECOVERY_WAIT_SECONDS value uses 30 seconds for OpenClaw gateway health and 90 seconds for Hermes gateway health.
Recreated-sandbox OpenShell registration uses 120 seconds when the recovery path does not supply another budget.
If the Ollama pull or post-create readiness timeout fires, onboarding emits the elapsed budget plus a hint to raise the relevant variable.
The Ollama pull preserves its partial download for the next attempt.
The ordinary post-create readiness wait deletes the orphaned sandbox first so the next nemo-deepagents onboard starts without that partially created sandbox.
A post-policy re-registration failure leaves the sandbox in place and reports that OpenShell did not re-register it.
Lifecycle Behavior Flags
The following flags change defaults for commands that manage existing sandboxes.
Legacy nemo-deepagents setup
Deprecated. Use nemo-deepagents onboard instead.
Running nemo-deepagents setup now delegates directly to nemo-deepagents onboard.