@vultr/model-provider-openclaw

Vultr Inference as a model provider plugin for OpenClaw. The model list, context windows, prices, input types and reasoning levels come from the live GET /v1/models catalog.

Install

openclaw plugins install clawhub:@vultr/model-provider-openclaw --accept-capabilities
export VULTR_INFERENCE_API_KEY=... # or: openclaw onboard --vultr-api-key <key>
openclaw models list --all --provider vultr --refresh
openclaw infer model run --local --model vultr/glm-5.3 --thinking high --prompt "hello"

To install from the Git repository instead:

openclaw plugins install git:github.com/vultr/model-provider-openclaw --force --accept-capabilities

--force is required for sources outside ClawHub. A local checkout installs the same way; OpenClaw copies it:

openclaw plugins install ~/src/vultr/model-provider-openclaw --force --accept-capabilities

For development, link the working tree into a throwaway profile instead of copying it:

npm install && npm run build
openclaw --profile vultr-dev plugins install --link . --force --accept-capabilities

How it works

OpenClaw loads dist/index.js, a single file built from index.ts with @vultr/model-catalog bundled in. It is committed: OpenClaw refuses a TypeScript entry outside a linked checkout, and it installs dependencies with --ignore-scripts, so nothing can be built on the user's machine.

index.ts registers provider vultr with three hooks:

• catalog.run feeds the model list and picker. Following OpenClaw's convention it returns null without an API key: listing needs no key, but a provider without one cannot be called
• prepareDynamicModel and resolveDynamicModel are what the agent runtime uses to resolve vultr/<id>. Without them a model that is in the list still fails with Unknown model. prepare loads the catalog; resolve is synchronous and answers from what prepare loaded

The last good catalog is kept in <agent dir>/cache/vultr-model-catalog.json and served when the network fails. With neither, the hooks return nothing and OpenClaw carries on without the provider.

Mapping

| OpenClaw | Catalog | | --- | --- | | contextWindow | contextWindow | | maxTokens | outputBudget: the smallest of maxOutputTokens, 65536 and a quarter of the context window | | cost (USD per million) | pricePerMillion: prompt, completion, cached prompt, cache write | | input | text, image, video, audio as accepted. The runtime model keeps text and image only | | reasoning, thinkingLevelMap | as in model-provider-pi | | compat | maxTokensField: "max_tokens", supportsDeveloperRole: false, supportsStore: false, supportsReasoningEffort, supportedReasoningEfforts, supportsTools |

The output budget exists because OpenClaw sends maxTokens as max_tokens without clamping it to the context that is left, and Vultr publishes max_length equal to the context window for most models. Unchanged, every request would ask for more than the engine can give and be rejected.

Environment

| Variable | Meaning | | --- | --- | | VULTR_INFERENCE_API_KEY | Detected by OpenClaw as the provider credential | | VULTR_INFERENCE_BASE_URL | Overrides https://api.vultrinference.com/v1 for the catalog and for requests |

Development

npm install
npm run typecheck
npm run build      # after any source change; commit dist/
npm test

Verifying against the real harness needs no API key: serve model-catalog-typescript/fixtures/vultr-catalog.json at /v1/models from a local server that records POST /v1/chat/completions and answers with a short SSE stream, then set VULTR_INFERENCE_BASE_URL to it and VULTR_INFERENCE_API_KEY to any value. Read the recorded request body: that is what Vultr would receive.