CoreWeave Serverless Inference — OpenClaw provider plugin

An OpenClaw model provider plugin for CoreWeave Serverless Inference (formerly Weights & Biases Inference) — open models (Kimi, GLM, DeepSeek, Qwen, Llama, Nemotron, gpt-oss, …) served on CoreWeave GPUs through an OpenAI-compatible API. Authenticate with your existing Weights & Biases account.

Provider id: coreweave · Model refs: coreweave/<vendor>/<model> (e.g. coreweave/moonshotai/Kimi-K2.6).

Install

openclaw plugins install clawhub:@coreweave/serverless-inference
openclaw gateway restart
openclaw plugins inspect coreweave --runtime --json

Setup

Get an API key at wandb.ai/authorize, then:

# Interactive
openclaw onboard --auth-choice coreweave-api-key

# Or via environment variable
export COREWEAVE_API_KEY="your-key"

# Non-interactive
openclaw onboard --non-interactive \
  --auth-choice coreweave-api-key \
  --coreweave-api-key "your-key"

Verify:

openclaw models list --provider coreweave
openclaw agent --model coreweave/moonshotai/Kimi-K2.6 --message "Hello, are you working?"

Project attribution (optional)

CoreWeave Serverless Inference can attribute usage to a specific W&B team/project via an openai-project header. This is optional — if you omit it, W&B uses your default entity and a project named inference. Set it only when you belong to more than one team or want usage attributed to a specific project:

{
  plugins: {
    entries: {
      coreweave: {
        config: { project: "my-team/my-project" },
      },
    },
  },
}

When project is set, OpenClaw attaches openai-project: team/project to every request (chat and model discovery). When unset, no header is sent.

Models

The plugin ships a verified catalog and filters it by IDs returned from the live /v1/models endpoint, falling back to the bundled catalog if discovery fails or returns no rows. Catalog metadata was checked on 2026-10-07 against the CoreWeave catalog feed. Models past their published retirement date are excluded. See catalog provenance for field mapping and lifecycle details.

Model ID	Context tokens	Input	Reasoning
moonshotai/Kimi-K2.7-Code	262144	text, image	Yes
moonshotai/Kimi-K2.6	262144	text, image	Yes
deepseek-ai/DeepSeek-V3.1	161000	text	No
Qwen/Qwen3.6-35B-A3B	262144	text, image	Yes
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B	262144	text	Yes
openai/gpt-oss-120b	131072	text	Yes
openai/gpt-oss-20b	131072	text	Yes
google/gemma-4-31B-it	262144	text, image	Yes
meta-llama/Llama-3.3-70B-Instruct	128000	text	No
meta-llama/Llama-3.1-8B-Instruct	131072	text	No
ibm-granite/granite-4.2-8b	131072	text	Yes
MiniMaxAI/MiniMax-M3	262144	text, image	Yes
zai-org/GLM-5.3-Flash	1048576	text, image	Yes
zai-org/GLM-5.2	1048576	text	Yes
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B	262144	text	Yes
Qwen/Qwen3.8-27B	262144	text, image	Yes
deepseek-ai/DeepSeek-V4-Pro-0813	1048576	text	Yes
deepseek-ai/DeepSeek-V4-Flash-0731	262144	text	Yes
deepseek-ai/DeepSeek-V4.1-Flash	1048576	text, image	Yes
google/gemma-4-26B-A4B-it	262144	text, image	Yes

openclaw models list --provider coreweave shows live model IDs that have verified metadata in this plugin. IDs missing from the verified catalog are skipped with a warning; update the plugin or configure their confirmed metadata explicitly. This prevents a stale discovery response from reintroducing a retired model.

Config file example

{
  env: { COREWEAVE_API_KEY: "..." },
  plugins: { entries: { coreweave: { config: { project: "my-team/my-project" } } } },
  agents: { defaults: { model: { primary: "coreweave/moonshotai/Kimi-K2.6" } } },
  models: {
    mode: "merge",
    providers: {
      coreweave: {
        baseUrl: "https://api.inference.wandb.ai/v1",
        apiKey: "${COREWEAVE_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "moonshotai/Kimi-K2.6",
            name: "Kimi K2.6",
            reasoning: true,
            input: ["text", "image"],
            cost: { input: 0.65, output: 3.41, cacheRead: 0.15 },
            contextWindow: 262144,
            maxTokens: 262144,
          },
        ],
      },
    },
  },
}

Troubleshooting

• API key not recognized — check that COREWEAVE_API_KEY is set without printing its value; rotate at wandb.ai/authorize.
• Attributing usage to a team/project — usage attribution is optional; set the project plugin config to team/project if you belong to multiple teams.
• Connection issues — the endpoint is https://api.inference.wandb.ai/v1; ensure your network allows HTTPS to that host.

Development

npm install        # installs openclaw (peer) + build/test tooling
npm run typecheck  # tsc --noEmit
npm test           # vitest
npm run build      # tsdown -> dist/index.js (+ .d.ts)

The published package ships built JavaScript: openclaw.runtimeExtensions points at ./dist/index.js, and prepublishOnly runs the build.

Publishing to ClawHub

Use the current ClawHub CLI from a reviewed checkout containing the intended release changes. Confirm the release version in package.json before building. From the repository root:

npm ci
npm run typecheck
npm test
npm run build
npm pack --dry-run
clawhub package validate .
clawhub package publish . --family code-plugin --owner coreweave --dry-run

The source argument . is the local package folder. ClawHub packs its contents into an npm tarball; the package name is read from package.json. Publishing to npm is not required for the clawhub: installation route. Review the dry-run output for the intended package name, version, GitHub repository, and exact source commit before publishing.

An account with publishing access to the coreweave ClawHub owner must perform the release. GitHub repository access does not grant that permission. From the same unchanged checkout:

clawhub login
clawhub whoami
clawhub package publish . --family code-plugin --owner coreweave --wait
clawhub package inspect @coreweave/serverless-inference --json

--wait waits for publication checks; an accepted upload alone does not mean the version is publicly installable. After publication, verify the reported version, then run the installation and setup steps above in a clean OpenClaw environment. Check model listing, streaming, and a tool-driven read/edit/reread on the supported runtime before declaring the release verified. See the ClawHub publishing requirements for ownership and release visibility details.

License

MIT — see MIT license.