NemoClawNemoClaw can route inference to a model server running on your machine instead of a cloud API. This page covers Ollama, compatible-endpoint paths for other servers, and experimental managed options for vLLM and NVIDIA NIM.
All approaches use the same inference.local routing model.
The agent inside the sandbox never connects to your model server directly.
OpenShell intercepts inference traffic and forwards it to the local endpoint you configure.
Ollama is the default local inference option. The onboard wizard detects Ollama automatically when it is installed or running on the host.
If Ollama is installed but not running, NemoClaw starts it for you.
On macOS and Linux, the wizard can also offer to install Ollama when it is not present.
When the host Ollama is below the minimum version NemoClaw expects for its starter models (currently 0.7.0), the wizard surfaces an explicit Upgrade Ollama entry in the provider menu instead of silently reusing the older daemon, and the express setup path resolves to that entry.
The wizard inspects both the CLI binary (ollama --version) and the locally running daemon (/api/version on :11434) so the upgrade entry still appears when only one side is stale, for example a fresh user-local binary paired with the original system daemon.
The gate skips Windows-host Ollama reached from WSL via host.docker.internal; the separate Use / Start / Install Ollama on Windows host entries handle that case and run their own actions on the Windows side.
On macOS, the wizard runs the platform install or upgrade path with brew upgrade ollama.
On Linux, the wizard runs the official https://ollama.com/install.sh path.
Upgrades on Linux always take the sudo-driven system path because the sudo-free user-local fallback would leave the existing system daemon on :11434 serving the stale binary.
If sudo is not available in a non-interactive run, NemoClaw refuses to silently downgrade the path and asks you to rerun interactively or upgrade Ollama manually.
After an upgrade finishes, NemoClaw re-probes the running daemon’s /api/version and fails the run if the daemon still reports below the minimum.
Fresh installs skip this re-probe because the bundled installers ship a daemon at or above the minimum.
On WSL, the wizard can use, start, restart, or install Ollama on the Windows host through PowerShell interop.
On native Linux, the install path picks between a system install (under /usr/local, via the official https://ollama.com/install.sh) and a sudo-free user-local install (under ${HOME}/.local).
NemoClaw selects the mode automatically:
sudo -n true returns 0) selects the system install.NEMOCLAW_NON_INTERACTIVE=1 or no TTY on stdin) without passwordless sudo selects the user-local install.
This is the path that lets headless hosts complete onboarding without prompting for a sudo password.Override the detection with NEMOCLAW_OLLAMA_INSTALL_MODE=system or NEMOCLAW_OLLAMA_INSTALL_MODE=user.
The user-local install replicates only the binary extraction step of the official installer.
It downloads the release tarball, extracts it to ${HOME}/.local, and launches ${HOME}/.local/bin/ollama serve once.
It does not configure a systemd service, does not create the ollama system user, and does not install CUDA drivers, so the daemon must be relaunched manually after a reboot.
NemoClaw also prints a one-line PATH hint if ${HOME}/.local/bin is not already on your PATH; you can add export PATH="${HOME}/.local/bin:$PATH" to your shell profile to invoke ollama directly.
Both modes rely on zstd for archive extraction. On Debian and Ubuntu, the system path uses sudo apt-get to install zstd automatically and explains the prompt before continuing.
The user-local path cannot bootstrap system packages without elevation, so if zstd is missing it prints per-distro install hints and exits — install zstd manually, then rerun onboarding.
Run the onboard wizard.
Select Local Ollama from the provider list.
NemoClaw lists installed models or offers starter models if none are installed.
On hosts where the larger starter models fit the currently available GPU memory, the starter list includes qwen3.6:35b and selects it by default.
When another GPU workload is using most of the memory at onboard time, NemoClaw downgrades the menu to the largest model that still fits.
It pulls the selected model, loads it into memory, and validates it before continuing.
When Ollama reports a loaded-model context length, NemoClaw uses that value for the contextWindow baked into openclaw.json unless you set NEMOCLAW_CONTEXT_WINDOW yourself.
If the selected model declares that it does not support tool calling, onboarding stops with guidance to choose a model whose ollama show <model> capabilities include tools.
The validation also requires structured chat-completions tool calls.
If the model leaks tool-call JSON as plain message text, onboarding stops so you can choose a model that returns tool calls in the expected response field.
If a host-side validation probe times out, NemoClaw retries the Ollama tool-call validation with a larger timeout before failing the setup.
On WSL, if you choose the Windows-host Ollama path, NemoClaw uses host.docker.internal:11434 and pulls missing models through the Ollama HTTP API instead of requiring the ollama CLI inside WSL.
When NemoClaw runs inside WSL, the provider menu can include Windows-host Ollama actions:
The install and restart paths set OLLAMA_HOST=0.0.0.0:11434 on the Windows side so Docker and WSL can reach the daemon through host.docker.internal.
After an install or restart action, NemoClaw relaunches Ollama from the detected Windows tray app or verified ollama.exe path and waits until host.docker.internal:11434 responds.
If the HTTP endpoint is not reachable yet, NemoClaw also checks for the Windows ollama.exe process through PowerShell interop so it can offer a start or restart action instead of hiding the Windows-host path.
If the daemon does not become reachable, onboarding prints PowerShell commands you can run to inspect the Windows-side process and port state. Use one Ollama instance on port 11434 at a time.
If both WSL and Windows-host Ollama are running, pick the intended menu entry during onboarding so NemoClaw validates and pulls models against the right daemon.
Windows-host Ollama requires Docker Desktop WSL integration because the sandbox reaches the Windows daemon through Docker Desktop’s WSL routing path. If NemoClaw detects native Docker Engine inside WSL, the provider menu labels Windows-host Ollama actions as requiring Docker Desktop integration. Selecting one of those actions in the unsupported native Docker topology exits early with a remediation message instead of trying to start or install Ollama on Windows.
Ollama is convenient for local chat, but some model/template combinations can
return tool calls as plain text under realistic agent load. If the TUI shows raw
JSON such as {"name":"memory_search","arguments":{...}} instead of running a
tool, switch to vLLM with --enable-auto-tool-choice and the correct
--tool-call-parser. See Tool-Calling Reliability.
On non-WSL hosts, NemoClaw keeps Ollama bound to 127.0.0.1:11434 and starts a token-gated reverse proxy on 0.0.0.0:11435.
The native install/start paths also reset NemoClaw-managed systemd launches to the loopback binding.
Containers and other hosts on the local network reach Ollama only through the
proxy, which validates a Bearer token before forwarding requests.
On that native path, NemoClaw never exposes Ollama without authentication.
WSL Ollama paths do not use this proxy.
Windows-host Ollama uses the Windows daemon through host.docker.internal.
For non-WSL Ollama setups, the onboard wizard manages the proxy automatically:
~/.nemoclaw/ollama-proxy-token with 0600 permissions.On native Linux hosts, a firewall can allow the host proxy health check while still blocking sandbox containers on the OpenShell Docker bridge. When the sandbox-side proxy probe fails with a TCP error, onboarding exits before it saves the inference route and prints a command like:
If the probe cannot run, for example because Docker Desktop or WSL uses a different host routing model, onboarding continues and relies on the regular proxy health check.
The sandbox provider is configured to use proxy port 11435 with the generated
token as its OPENAI_API_KEY credential.
OpenShell’s L7 proxy injects the token at egress, so the agent inside the
sandbox never sees the token directly.
All proxy endpoints require the Bearer token, including GET /api/tags.
Internal health and reachability checks run via the proxy treat any HTTP
response (including 401) as proof the proxy is alive — they only fail
when nothing answers at all.
If Ollama is already running on a non-loopback address when you start onboard,
the wizard restarts it on 127.0.0.1:11434 so the proxy is the only network
path to the model server.
When you switch away from Ollama, stop host services, or destroy an Ollama-backed sandbox, NemoClaw asks Ollama to unload currently loaded models from GPU memory.
The cleanup sends keep_alive: 0 for each model reported by Ollama and runs on a best-effort basis, so shutdown continues if Ollama is already stopped.
This does not delete downloaded model files.
If NEMOCLAW_MODEL is not set, NemoClaw selects a default model based on available memory.
If NEMOCLAW_MODEL names a known bootstrap model (for example qwen3.6:35b) that does not fit the host’s currently available GPU memory, NemoClaw warns and falls back to the largest known model that does fit.
Unknown or custom tags (any value the bootstrap registry has not seen) are still passed through; the Ollama runner validates the choice itself.
--yes (or NEMOCLAW_YES=1) authorises the Ollama model download without an interactive confirmation prompt.
Under --non-interactive, --yes (or NEMOCLAW_YES=1) is required to authorise the download — onboard exits otherwise, since it cannot prompt.
Run onboard without --non-interactive to get the interactive [y/N] prompt that shows the model size before downloading.
This option works with any server that implements /v1/chat/completions, including vLLM, TensorRT-LLM, llama.cpp, LocalAI, and others.
For compatible endpoints, NemoClaw uses /v1/chat/completions by default.
This avoids a class of failures where local backends accept /v1/responses requests but silently drop the system prompt and tool definitions.
To opt in to /v1/responses, set NEMOCLAW_PREFERRED_API=openai-responses before running onboard.
Start your model server. The examples below use vLLM, but any OpenAI-compatible server works.
Run the onboard wizard.
When the wizard asks you to choose an inference provider, select Other OpenAI-compatible endpoint.
Enter the base URL of your local server, for example http://localhost:8000/v1.
The wizard prompts for an API key.
If your server does not require authentication, enter any non-empty string (for example, dummy).
NemoClaw validates the endpoint by sending a test inference request before continuing.
The wizard probes /v1/chat/completions by default for the compatible-endpoint provider.
If you set NEMOCLAW_PREFERRED_API=openai-responses, NemoClaw probes /v1/responses instead and only selects it when the response includes the streaming events OpenClaw requires.
If a reasoning model returns only reasoning content before producing a final answer, NemoClaw retries the smoke request with a larger response budget.
Route, configuration, and authentication failures still fail immediately.
Set the following environment variables for scripted or CI/CD deployments.
For the compatible-endpoint provider, /v1/chat/completions is the default.
NemoClaw tests streaming events during onboarding and uses chat completions
without probing the Responses API.
To opt in to /v1/responses, set NEMOCLAW_PREFERRED_API before running onboard:
The wizard then probes /v1/responses and only selects it when streaming
support is complete.
If the probe fails, the wizard falls back to /v1/chat/completions
automatically.
You can use this variable in both interactive and non-interactive mode.
If you already onboarded and the sandbox is failing at runtime, re-run nemoclaw onboard to re-probe the endpoint and bake the correct API path
into the image.
Refer to Switch Inference Models for more information.
If your local server implements the Anthropic Messages API (/v1/messages), choose Other Anthropic-compatible endpoint during onboarding instead.
For non-interactive setup, use NEMOCLAW_PROVIDER=anthropicCompatible and set COMPATIBLE_ANTHROPIC_API_KEY.
When vLLM is already running on localhost:8000, NemoClaw can detect it automatically and query the /v1/models endpoint to determine the loaded model.
On supported Linux hosts with NVIDIA GPUs, the onboard wizard can also install or start a managed vLLM container for you.
For an already-running vLLM server, run nemoclaw onboard and select Local vLLM [experimental] from the provider list.
If vLLM is already running, NemoClaw detects the running model and validates the endpoint.
If vLLM is not running and your host matches a DGX Spark or DGX Station managed profile, NemoClaw shows the Install vLLM or Start vLLM entry by default.
Generic Linux NVIDIA GPU hosts still require NEMOCLAW_EXPERIMENTAL=1 or NEMOCLAW_PROVIDER=install-vllm before the managed entry appears.
NemoClaw pulls the vLLM image, downloads model weights into ~/.cache/huggingface, starts the nemoclaw-vllm container on localhost:8000, and prints progress markers while the model loads.
The first run can take 10 to 30 minutes.
Later runs reuse the cached image and model weights.
Managed vLLM uses these profiles:
NemoClaw forces the chat/completions API path for vLLM.
The vLLM /v1/responses endpoint does not run the --tool-call-parser, so tool calls arrive as raw text.
Use an already-running vLLM server:
Install or start managed vLLM when a supported profile is detected.
On DGX Spark and DGX Station, NEMOCLAW_PROVIDER=install-vllm is enough for non-interactive runs; add NEMOCLAW_EXPERIMENTAL=1 on generic Linux NVIDIA GPU hosts.
NemoClaw records the model returned by vLLM’s /v1/models endpoint.
Start vLLM with the model you want before onboarding if you manage the server yourself.
Managed vLLM serves the profile default unless you select a different registry entry.
Export NEMOCLAW_VLLM_MODEL=<slug> before invoking the installer to choose a different model from the registry.
NemoClaw uses the matching vllm serve flags, including the reasoning parser, tool-call parser, and --max-model-len.
Recognised slugs:
The slug is case-insensitive; the full Hugging Face id is also accepted. An unrecognised value fails fast with a list of valid slugs.
Gated models require a Hugging Face token; export it before onboarding so NemoClaw can forward it into the managed vLLM container:
HUGGING_FACE_HUB_TOKEN is accepted as an alternative.
The token check runs on the host before any docker pull, so a missing or empty token aborts onboarding before bandwidth is spent on a 401.
NemoClaw can pull, start, and manage a NIM container on hosts with a NIM-capable NVIDIA GPU.
Set the experimental flag and run onboard.
Select Local NVIDIA NIM [experimental] from the provider list. NemoClaw filters available models by GPU VRAM, pulls the NIM container image, starts it, and waits for it to become healthy before continuing. On hosts with mixed NVIDIA GPU models, the preflight summary shows each detected GPU model and the total VRAM so you can confirm which device class the model selection used.
NIM container images are hosted on nvcr.io and require NGC registry authentication before docker pull succeeds.
If Docker is not already logged in to nvcr.io, onboard prompts for an NGC API key and runs docker login nvcr.io over --password-stdin so the key is never written to disk or shell history.
The prompt masks the key during input and retries once on a bad key before failing.
In non-interactive mode, onboard exits with login instructions if Docker is not already authenticated; run docker login nvcr.io yourself, then re-run nemoclaw onboard --non-interactive.
If NGC_API_KEY or NVIDIA_API_KEY is already exported, NemoClaw passes it into the managed NIM container through the process environment instead of command-line arguments.
If the NIM container exits before the health endpoint becomes ready, onboarding stops early and prints the last container log lines.
NIM uses vLLM internally.
The same chat/completions API path restriction applies.
To select a specific model, set NEMOCLAW_MODEL.
Local inference requests use a default timeout of 180 seconds. Large prompts on hardware such as DGX Spark can exceed shorter timeouts, so NemoClaw sets a higher default for Ollama, vLLM, NIM, and compatible-endpoint setup.
To override the timeout, set the NEMOCLAW_LOCAL_INFERENCE_TIMEOUT environment variable before onboarding:
The value is in seconds. This setting is baked into the sandbox at build time.
Changing it after onboarding requires re-running nemoclaw onboard.
NEMOCLAW_LOCAL_INFERENCE_TIMEOUT only governs the inference-server validation probe.
During local Ollama setup, NemoClaw treats host-side curl process timeouts as retryable probe failures and retries with a larger timeout before it reports a validation failure.
NemoClaw also retries Docker runtime detection with a longer docker info timeout before it chooses the local inference route.
The post-create readiness wait (image build, gateway upload, in-sandbox boot) has its own budget, NEMOCLAW_SANDBOX_READY_TIMEOUT, also defaulting to 180 seconds.
On hosts where the sandbox image takes minutes to build or upload — large quantised models, DGX Station first runs, or remote VMs over a slow link — raise both together:
If onboard ends with Sandbox '<name>' was created but did not become ready within 180s, refer to Troubleshooting.
After onboarding completes, confirm the active provider and model.
The output shows the provider label (for example, “Local vLLM” or “Other OpenAI-compatible endpoint”) and the active model.
For Local Ollama, status also checks the authenticated proxy when a proxy token is available.
If Inference is healthy but Inference (auth proxy) is not, rerun onboarding to repair the proxy path that sandbox requests use.
You can change the model without re-running onboard. Refer to Switch Inference Models for the full procedure.
For compatible endpoints, the command is:
If the provider itself needs to change (for example, switching from vLLM to a cloud API), pass the new provider to nemoclaw inference set.