Featherless AI Provider Setup and Model Selection

Featherless AI exposes open models via an API that follows OpenAI conventions. OpenClaw ships Featherless as an officially supported external provider plugin, which keeps the default catalog compact while still accepting any exact model id Featherless offers at runtime.

Setup

After installing the plugin, restart the Gateway:

openclaw plugins install @openclaw/featherless-provider
openclaw gateway restart

To begin onboarding, run:

openclaw onboard --auth-choice featherless-api-key

For setups that skip interactive prompts:

openclaw onboard --non-interactive --accept-risk --skip-health \
  --mode local \
  --auth-choice featherless-api-key \
  --featherless-api-key "$FEATHERLESS_API_KEY"

Alternatively, pass the key to the Gateway process through the environment:

export FEATHERLESS_API_KEY="<your-featherless-api-key>" # pragma: allowlist secret

To confirm the provider is working:

openclaw models list --provider featherless

Default model

Because Featherless's documentation describes native tool calling for the Qwen 3 family, the plugin defaults to Qwen/Qwen3-32B during setup. OpenClaw sets a 32,768-token context window, a cautious 4,096-token cap on output, and Qwen chat-template thinking controls.

Pricing fields in the catalog draw on Featherless's published per-request rates: $0.102 per million input tokens and $0.493 per million output tokens. Fixed subscription plans are charged at a flat rate, and cache cost fields remain at zero since Featherless does not list separate cache-read or cache-write pricing for this model.

Other Featherless models

Place the exact Featherless model id after the featherless/ provider prefix:

{
  agents: {
    defaults: {
      model: {
        primary: "featherless/moonshotai/Kimi-K2-Instruct",
      },
    },
  },
}

OpenClaw intentionally avoids pulling Featherless's entire public model list into the picker. That index is extensive but lacks enough structured capability metadata to safely categorize every text, vision, embedding, and reasoning model. As a result, unknown ids fall back to conservative text-only, non-reasoning defaults: a 4,096-token context window and a 1,024-token output limit.

When a model requires different metadata, add an explicit provider model entry:

{
  models: {
    mode: "merge",
    providers: {
      featherless: {
        baseUrl: "https://api.featherless.ai/v1",
        apiKey: "${FEATHERLESS_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "google/gemma-3-27b-it",
            name: "Gemma 3 27B",
            input: ["text", "image"],
            reasoning: false,
            contextWindow: 32768,
            maxTokens: 4096,
          },
        ],
      },
    },
  },
}

Before adding custom metadata, review Featherless's model catalog to confirm current availability and capability tags.

Troubleshooting

• 401 or 403: verify FEATHERLESS_API_KEY is reachable by the Gateway process, or rerun onboarding.
• Unknown model: after the featherless/ prefix, use the exact case-sensitive id from Featherless.
• Tool calls appearing as text: pick a model family Featherless lists for native function calling, for instance Qwen 3.
• Managed Gateway missing the key: place it in ~/.openclaw/.env or another environment source the service loads, then restart the Gateway.

Related

• Model providers
• All providers
• Thinking modes