Use ModelStream with OpenClaw

OpenClaw supports any OpenAI-compatible provider through its models.providers config. ModelStream exposes a single OpenAI-compatible endpoint in front of 300+ models, so you register it once as a custom provider and then switch models by changing one string — no new auth, no code changes.

Skip this if OpenClaw is already installed.

curl -fsSL https://openclaw.ai/install.sh | bash

iwr -useb https://openclaw.ai/install.ps1 | iex

The installer detects your OS, installs Node if needed (Node 24 recommended, 22.19+ minimum), and launches onboarding. Prefer managing Node yourself? npm install -g openclaw@latest && openclaw onboard --install-daemon.

Verify:

openclaw --version
openclaw gateway status

1. 1Sign up at modelstream.ai.
2. 2Go to Dashboard → API Keys → Create Key, name it (e.g. openclaw), and copy it.
3. 3Export it where the OpenClaw gateway runs:

export MODELSTREAM_API_KEY="ms-..."

Add this to your OpenClaw config (~/.openclaw/openclaw.json):

{
  agents: {
    defaults: {
      model: { primary: "modelstream/gpt-4o-mini" },
    },
  },
  models: {
    mode: "merge",
    providers: {
      modelstream: {
        baseUrl: "https://api.modelstream.ai/v1",
        apiKey: "${MODELSTREAM_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "gpt-4o-mini",
            name: "GPT-4o mini (via ModelStream)",
            input: ["text", "image"],
            contextWindow: 128000,
            maxTokens: 16384,
          },
          {
            id: "claude-sonnet",
            name: "Claude Sonnet (via ModelStream)",
            input: ["text", "image"],
            contextWindow: 200000,
            maxTokens: 64000,
          },
          {
            id: "modelstream-auto",
            name: "ModelStream Smart Routing",
            contextWindow: 128000,
            maxTokens: 16384,
          },
        ],
      },
    },
  },
}

Notes that save you a support ticket:

• mode: "merge" keeps OpenClaw's built-in provider catalog alongside ModelStream.
• Every model you want to call must appear in the provider's models[] list. If you also set agents.defaults.models (it acts as an allowlist), add each modelstream/<id> there too — otherwise you'll hit the classic "model not allowed" error even with a valid key.
• Model refs are provider/model, e.g. modelstream/claude-sonnet.
• Set input: ["text", "image"] on vision-capable models so image attachments are passed natively.
• contextWindow/maxTokens are optional (defaults: 200000 / 8192) but set real values per model.

openclaw models set modelstream/gpt-4o-mini
openclaw models list     # ModelStream models should appear
openclaw doctor          # config sanity check

Then send your agent a message and run /models in the chat — your ModelStream models should be listed and selectable.

Switch models in one line. Routing the same agent to a different model is just:

openclaw models set modelstream/claude-sonnet

Or let ModelStream pick the cheapest model that meets your quality bar:

openclaw models set modelstream/modelstream-auto

ModelStream already does multi-region provider failover behind one model slug. For model-level fallbacks inside OpenClaw (e.g. fall back to a cheaper model if your primary is rate-limited), configure a fallback chain — see Model failover in the OpenClaw docs.

• Browse the full catalog in your ModelStream dashboard and add slugs to the config as you need them.
• Estimate your monthly bill with the ModelStream cost calculator.
• Coming from OpenAI or Anthropic elsewhere in your stack? The migration is the same one-line base-URL swap — see the ModelStream migration guides.