Network Policies
NemoClaw runs with a deny-by-default network policy. The sandbox can only reach endpoints that are explicitly allowed. OpenShell intercepts any request to an unlisted destination and prompts the operator to approve or deny it in real time through the TUI.
Baseline Policy
Deep Agents sandboxes use an agent-specific baseline policy in agents/langchain-deepagents-code/policy-additions.yaml so the managed dcode runtime can reach inference, GitHub, and package endpoints while keeping the same deny-by-default model. Deep Agents uses strict Landlock compatibility, so sandbox startup fails when OpenShell cannot enforce the managed filesystem policy.
Filesystem
/dev/pts is the pseudo-terminal (devpts) directory. It is writable so PTY-based tools (tmux, script, and interactive shells) can allocate a terminal. Without it, those tools fail with fork failed: Permission denied.
Read-only access to /var/lib/dpkg lets dpkg-query inspect installed package metadata. The filesystem policy does not grant write access to the package database.
The sandbox process runs as a dedicated sandbox user and group.
For Deep Agents, Landlock enforcement is strict. If the kernel or runtime cannot enforce the managed filesystem policy, sandbox startup fails closed.
Network Policies
The following endpoint groups are allowed by default:
The separate raw.githubusercontent.com route lets Deep Agents Code follow GitHub file links and read repository source through its managed fetch_url tool. Repository, ref, and file path segments vary by task, so the route covers the host while limiting requests to read-only GET and HEAD methods and the listed managed binaries.
All endpoints use TLS termination and are enforced at port 443.
Deep Agents is a terminal-oriented harness. NemoClaw does not configure messaging channel bridges
for Deep Agents today, and the baseline policy does not include Tavily, LangSmith, MCP hosts, or
arbitrary web endpoints. Apply maintained presets such as github, pypi, or tavily only when
the sandbox needs that access.
Policy Tiers
During onboarding, the wizard prompts for a policy tier that determines the default set of presets applied on top of the baseline policy. The baseline policy is always applied regardless of the selected tier. An operator can exclude a specific baseline entry from the current OpenShell policy with policy exclude when they accept reduced, unsupported functionality. NemoClaw stores no separate exclusion record. Rebuild and clone carry the current OpenShell policy forward, so changes made through NemoClaw, the OpenShell TUI, or direct host-side policy editing have the same lifecycle.
Balanced uses brew-balanced to allow only GET and HEAD to raw.githubusercontent.com for the preset’s binaries.
Open retains unrestricted raw GitHub access through brew, and Personal retains its broad web policy.
Reselecting Balanced’s suggested presets preserves an existing brew grant, including its unrestricted raw GitHub access.
Adding brew-balanced does not narrow that broader grant.
To narrow it, explicitly remove brew and apply brew-balanced.
Resuming onboarding reapplies the saved selection; skipping optional setup preserves existing grants.
When Personal is selected or carried forward, the personal-open-internet preset is mandatory for every agent and every onboarding entry point. Interactive choices, NEMOCLAW_POLICY_MODE=custom, and NEMOCLAW_POLICY_MODE=skip control only additional presets; they cannot deselect, skip, or replace Personal’s required web authority.
The Personal tier applies the personal-open-internet policy preset with a hostless L4 endpoint on destination ports 80 and 443.
The policy matches any requested host on either port, then permits the connection only when every resolved address is in the preset’s allowed ranges.
The rule does not inspect the application protocol or payload, so traffic on these ports is not limited to HTTP or HTTPS.
OpenShell does not restrict the hostname, HTTP method, path, or body after the rule permits the connection.
An agent can send workspace data or sandbox-visible credentials to an arbitrary reachable service on either port without an operator approval prompt.
The preset excludes unspecified, loopback, and link-local address ranges, including the common cloud metadata range. OpenShell also keeps its hard blocks for those destinations. Other destination ports remain denied unless another policy entry permits them. The sandbox’s filesystem, process, gateway authentication, and managed credential controls remain active. Use this tier only for trusted personal workloads with trusted prompts and data.
After selecting a tier, a combined preset and access-mode screen lets you include or exclude optional presets and toggle each between read (GET only) and read-write (GET + POST/PUT/PATCH) access. On Personal, NemoClaw restores personal-open-internet if it is deselected in the screen. Tier-default presets are pre-selected; additional presets can be added from the built-in preset list available to the sandbox’s active agent. NemoClaw filters tier defaults and built-in preset choices by the active agent’s supported integrations. The personal-open-internet preset uses L4 passthrough, so its read-write label does not add HTTP method or path inspection.
Deep Agents can use the maintained tavily opt-in path, but messaging channel presets are omitted because the terminal harness does not run a NemoClaw messaging bridge today.
NemoClaw automatically suggests the preset that matches the selected provider and removes stale web search presets during resume reconciliation when you switch providers or disable web search.
The Personal tier does not select a Brave Search or Tavily Search preset by default.
Its broad route supports ordinary keyless web fetches, while web_search still requires a separately configured provider.
Explicit custom preset lists and manual interactive selections remain operator-controlled for additional presets.
When LangChain Deep Agents Code is onboarded with --observability, NemoClaw adds the observability-otlp-local preset on Balanced, Open, and Personal tiers.
The Restricted tier suppresses this agent-required preset during onboarding and rebuild reconciliation.
An operator can add it manually after reviewing the additional egress, but the next Restricted reconciliation removes it.
The applied set therefore reflects the chosen tier plus any agent-required presets, so policy list may show one or more presets that do not appear in the tier table above.
The policy list provenance tags are inferred from the current tier YAML and the active agent at display time and are not persisted per preset.
A preset whose name matches an entry in the sandbox’s current tier definition 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.
Claude Code direct egress is not included in the Restricted, Balanced, or Open tiers.
Personal’s broad route permits its port 443 transport but does not install or configure Claude Code; on other tiers, apply the claude-code preset explicitly if you install and run the CLI inside the sandbox with its own credentials.
Normal NemoClaw Anthropic inference still routes through the OpenShell gateway.
Tier definitions are stored in nemoclaw-blueprint/policies/tiers.yaml.
In non-interactive mode, set the tier with NEMOCLAW_POLICY_TIER:
Unset, blank, or whitespace-only NEMOCLAW_POLICY_TIER values use the balanced default. In non-interactive onboarding, a non-blank value that does not match a known tier exits before preflight, gateway, or inference side effects and lists the valid options. Interactive onboarding ignores an invalid environment value and shows the normal tier prompt.
Inference
The baseline policy allows only the local inference route. External inference providers are reached through the OpenShell gateway, not by direct sandbox egress.
Local OTLP Trace Export
The observability-otlp-local preset supports the opt-in LangChain Deep Agents Code trace path. It is not a general remote observability policy.
The sandbox sends OTLP/HTTP protobuf traces to a collector that the operator runs on the host. The managed exporter uses standard OTLP transport headers but does not accept operator-supplied custom or authentication headers. Remote backend endpoints and credentials stay in that collector. The policy does not allow direct LangSmith, Jaeger, Phoenix, or other backend egress from Deep Agents Code. Changing the collector’s downstream exporter requires no sandbox policy change.
OpenShell observes Deep Agents Code export as the managed Python interpreter, so this permission is process-wide for /opt/venv/bin/python3* rather than limited to the dcode launcher. Sandbox Python can forge spans and resource attributes. The explicit --observability opt-in can export bounded prompts, responses, tool arguments, tool results, and operational metadata. Managed size and recognized-key redaction do not detect secrets embedded in ordinary content values. The collector must enforce the operator’s filtering and redaction requirements before forwarding traces, and it must not treat span fields such as service.name as authenticated tenant identity. For a safe host binding, policy recovery commands, and a runnable collector, refer to Set Up Deep Agents Trace Export.
Operator Approval Flow
When the agent attempts to reach an endpoint not listed in the policy, OpenShell intercepts the request and presents it in the TUI for operator review. The Personal tier does not prompt for matching TCP connections on destination ports 80 or 443 because personal-open-internet already permits them. The flow has these steps:
- The agent makes a network request to an unlisted host.
- OpenShell blocks the connection and logs the attempt.
- The TUI command
openshell termdisplays the blocked request with host, port, and requesting binary. - The operator approves or denies the request.
- If approved, the endpoint is added to the running policy for the session.
To monitor requests as the agent runs, open the TUI:
For step-by-step navigation and approval controls, refer to Approve or Deny Agent Network Requests.
Modifying the Policy
Static Changes
Edit agents/langchain-deepagents-code/policy-additions.yaml and re-run the onboard wizard:
Dynamic Changes
Apply policy updates to a running sandbox without restarting:
To replace the live policy with a complete base policy file, export the current base policy and use openshell policy set.
This workflow requires OpenShell 0.0.72+ for policy get --base and policy set --wait.
NemoClaw parses the policy, strips the OpenShell metadata header, and replaces literal credentials with [STRIPPED_BY_MIGRATION]. Replace every marker with a supported OpenShell credential binding or resolver placeholder before you apply the file.
NemoClaw exits nonzero if it cannot validate the base policy.
Do not add --raw when you plan to edit and reapply the file.
Edit or review current-policy.yaml, then apply it:
Excluding a baseline entry
The baseline applies to every sandbox, but an operator can remove one exact baseline entry from the current OpenShell policy after accepting the reduced-support impact.
Preview and apply the live change:
The command prints the endpoints, method and path rules, binaries, and supported features affected by the removal. It refuses entries without a reviewed feature-impact disclosure and refuses keys required by an active preset. The critical managed_inference entry cannot be excluded because it carries the required managed-inference route.
NemoClaw reads the current OpenShell policy, removes the selected key, writes the complete modified document, and verifies the live result. It does not create an exclusion record, journal, or replay state. Rebuild and clone preserve the current OpenShell policy as a whole, so a change made with this command has the same lifecycle as one made through the OpenShell TUI or another trusted host process.
Restore an entry from the current agent baseline with:
If the current baseline still defines the key, restore previews and then adds that current entry to the live OpenShell policy. If the baseline no longer defines it, the command reports that there is nothing to restore and leaves the live policy unchanged. Both mutation commands require acknowledgement unless --dry-run is used. A non-interactive run requires --force, --yes, or -y.
Excluding a baseline entry leaves agent features that depend on it unsupported for that sandbox.