OpenClaw provider plugin for Crusoe: open-weight models on Crusoe's GPU cloud.
openclaw-plugin-crusoe
An OpenClaw provider plugin for Crusoe Managed Inference. It adds Crusoe to OpenClaw's model setup, so an OpenClaw assistant can run on the open-weight models Crusoe serves behind https://api.inference.crusoecloud.com/v1.
OpenClaw routes third-party providers through plugins that the vendor publishes and maintains. This folder is the source of that plugin.
| Property | Value |
|---|---|
| Provider id | crusoe |
| Auth env var | CRUSOE_API_KEY |
| Onboarding flag | --auth-choice crusoe-api-key |
| Direct CLI flag | --crusoe-api-key <key> |
| API | OpenAI-compatible (openai-completions) |
| Base URL | https://api.inference.crusoecloud.com/v1 |
| Default model | crusoe/openai/gpt-oss-120b |
Prerequisites
- OpenClaw 2026.11.3 or later
- Node.js 24.16 or later on the 24 line, or 26.1 or later, to build the plugin
- A Crusoe Inference API key from the Crusoe Console (Intelligence Foundry > Inference > Create API Key)
Install
Build the plugin from this folder and install the archive:
cd integrations/openclaw
npm install
npm run build
npm pack
openclaw plugins install ./openclaw-plugin-crusoe-0.1.0.tgz --force --accept-capabilities
--force confirms an install from a local archive, and --accept-capabilities accepts the provider capability the plugin declares.
Once the package is published on ClawHub, the install is one command:
openclaw plugins install openclaw-plugin-crusoe
Setup
Give OpenClaw the key:
openclaw onboard --auth-choice crusoe-api-key
For scripted installs:
openclaw onboard --non-interactive --accept-risk --skip-health \
--auth-choice crusoe-api-key \
--crusoe-api-key "$CRUSOE_API_KEY"
Check that the models are available:
openclaw models list --provider crusoe --refresh
Models
With a key saved, the plugin lists every public chat model the key can see from GET /v1/models and keeps Crusoe's ids, context windows, and prices. A new Crusoe model needs no plugin release: pick it as crusoe/<model id> once it appears in the list. Private dedicated deployments and non-chat rows such as embeddings are not listed as chat models.
These models carry reviewed metadata:
| Model ref | Input | Context | Thinking |
|---|---|---|---|
crusoe/openai/gpt-oss-120b | text | 131,072 | low/medium/high (default medium) |
crusoe/nvidia/Nemotron-3-Nano-Omni-Reasoning-30B-A3B | text, image | 262,144 | off/on (default on) |
crusoe/google/gemma-4-31b-it | text, image | 262,144 | off/on (default off) |
If the model list cannot be read, the plugin serves these three rows. Check the available models page for the current catalog.
Thinking
Crusoe models control reasoning through the OpenAI reasoning_effort field.
- GPT-OSS 120B accepts
low,medium, andhigh, and each level changes how long the model reasons. It cannot turn reasoning off. - Nemotron 3 Nano Omni and Gemma 4 are hybrid-reasoning models:
noneturns reasoning off and any other level turns it on. The picker offersoffandon.
Models that appear only in the live list are sent no reasoning_effort and reason the way the endpoint defaults them. Configure a custom model row if one needs a specific ladder.
Custom base URL
Setting models.providers.crusoe.baseUrl routes chat requests to that URL. The plugin then serves the reviewed rows and does not call the model list, so the key is listed only against the Crusoe host.
The URL must use https. Plain http is accepted only for localhost, 127.0.0.1 and [::1]; with any other http URL the plugin offers no catalog, so the key is never paired with an unencrypted endpoint.
Develop
npm ci
npm run build
npm test
npm run validate
npm run validate builds the plugin and runs clawhub package validate . --out .clawhub-validation, the same inspection ClawHub applies on publish.
The build and the tests do not install OpenClaw, so the lockfile holds only the build, test and validation tools. src/openclaw-plugin-sdk.d.ts declares the parts of openclaw/plugin-sdk that the plugin uses, and test/sdk/ stands in for the two SDK functions it calls. npm run validate is the check against a real release: it inspects the plugin against the latest published OpenClaw. When the plugin starts using more of the SDK, or openclaw.build.openclawVersion moves, update the declarations and the stand-ins to match.
The reviewed rows live in src/models.ts and are mirrored in openclaw.plugin.json. A test fails when they drift.
Publish
Releases go to ClawHub from a maintainer's machine:
npm exec clawhub -- login
npm run validate
npm exec clawhub -- package publish . --dry-run
npm exec clawhub -- package publish . --wait
Bump version in package.json and openclaw.plugin.json together before each release.
