Switching Model Providers in NemoClaw

Step 1: see what you're on now

nemoclaw inference get
# Provider: nvidia-prod
# Model:    nvidia/nemotron-3-super-120b-a12b

Nemotron 3 Super is the default for OpenClaw and Hermes sandboxes on NVIDIA Endpoints. Add --json for scripts.

Step 2: make sure the provider is registered

nemoclaw credentials list

Canonical provider IDs:

If the provider you want isn't listed, run nemoclaw onboard and choose it. The wizard checks the key and model before saving anything. Some installer-style aliases such as build and custom are also accepted.

Step 3: switch, with provider and model together

# Claude
nemoclaw inference set --provider anthropic-prod --model claude-sonnet-5 --sandbox my-assistant

# OpenAI
nemoclaw inference set --provider openai-api --model <openai-model-id> --sandbox my-assistant

# Back to Nemotron
nemoclaw inference set --provider nvidia-prod --model nvidia/nemotron-3-super-120b-a12b --sandbox my-assistant

Changing only the model on the same provider uses the same command; the provider is still required. What happens next:

• OpenClaw sandboxes: NemoClaw updates the model in the running config and restarts only the OpenClaw gateway. Expect a brief interruption. If the restart fails, the change is still saved; run nemoclaw my-assistant gateway restart. If the agent then won't answer, run nemoclaw my-assistant doctor --fix.
• Hermes sandboxes: NemoClaw rewrites /sandbox/.hermes/config.yaml (model, base URL, API type and the model's context length) without restarting Hermes.
• Local and custom endpoints are checked from inside the sandbox before the switch is kept. If the check fails, NemoClaw tries to put the previous route back.

The model IDs in NVIDIA's onboarding menu lag behind the providers. The menu still lists claude-sonnet-4-6 and gpt-5.4, but inference set accepts any model ID the provider serves and validates it. Current prices are in the cost calculator.

Your own server or gateway (compatible endpoints)

For vLLM, SGLang, a self-hosted NIM, LiteLLM or a corporate gateway, supply the complete route in one go:

export COMPATIBLE_API_KEY="<api-key>"
nemoclaw inference set \
  --provider compatible-endpoint \
  --model <model-name> \
  --endpoint-url <trusted-url> \
  --credential-env COMPATIBLE_API_KEY \
  --inference-api openai-completions \
  --sandbox my-assistant

--inference-api takes openai-completions, openai-responses or anthropic-messages. To change an existing compatible provider's URL or key, run nemoclaw onboard again; inference set won't overwrite that binding. NemoClaw's support covers its adapter, not every compatible proxy, so test tool calling before relying on it.

Several sandboxes on one gateway

One OpenShell gateway has one live route. Every sandbox registered on that gateway uses the same provider and model. inference set names the other sandboxes it would affect, warns you before switching, and refuses a change that would leave another sandbox with a different recorded route. To run different models side by side, give each sandbox its own gateway.

Step 4: verify

nemoclaw inference get                 # route shows the new provider and model
nemoclaw my-assistant status           # sandbox healthy

Then send the agent a test message from its chat channel or nemoclaw my-assistant connect. If a switch reports that rollback couldn't complete, run nemoclaw onboard before trying again.

Official references: NemoClaw docs Inference section. Related: local GPU inference, setup.