openclaw/docs/providers/openrouter.md

summary, read_when, title

summary	read_when	title
Use OpenRouter's unified API to access many models in OpenClaw	You want a single API key for many LLMs You want to run models via OpenRouter in OpenClaw You want to use OpenRouter for image generation You want to use OpenRouter for music generation You want to use OpenRouter for video generation	OpenRouter

OpenRouter routes requests to many models behind one API and one key. It is OpenAI-compatible, so OpenClaw talks to it over the same openai-completions-style transport used for other proxy providers.

Getting started

In a private chat, send /login openrouter or select OpenRouter from /login. Choose Sign in with OpenRouter, approve access in your browser, and return to chat. OpenClaw receives the browser callback and saves the credential before reporting success. Use /login cancel to cancel a pending sign-in.

Login saves access without choosing a starter model. If current model restrictions hide OpenRouter models, choose Show all OpenRouter models or Keep current restrictions. The credential stays saved either way. In the Control UI, use Settings → Models → Connect for the same credential-only flow, then use the model menu to choose a model from the Gateway's catalog.

Chat browser sign-in uses the Gateway's managed Tailscale HTTPS address. With Tailscale Serve, your browser must have access to the same tailnet. If no managed HTTPS address is available, enable Serve and retry, or use the CLI flow below. The callback does not sign you in to the Control UI.

The Control UI receives the return automatically at the managed HTTPS address or when opened directly on the local Gateway's loopback address and port. Other addresses use manual redirect completion. If pasted input is incomplete or invalid, correct it in the same sign-in attempt and submit again.

```bash openclaw onboard --auth-choice openrouter-oauth ```

    OpenClaw opens OpenRouter's browser sign-in flow (PKCE), exchanges the
    code for an OpenRouter API key, and stores it in the default
    OpenRouter auth profile. On remote/headless hosts, OpenClaw prints the
    sign-in URL and asks you to paste the redirect URL after signing in.
  </Step>
  <Step title="(Optional) Switch to a specific model">
    Onboarding defaults to `openrouter/auto`. Pick a concrete model later:

    ```bash
    openclaw models set openrouter/<provider>/<model>
    ```

  </Step>
</Steps>

Create an API key at [openrouter.ai/keys](https://openrouter.ai/keys). ```bash openclaw onboard --auth-choice openrouter-api-key ``` Onboarding defaults to `openrouter/auto`. Pick a concrete model later:

    ```bash
    openclaw models set openrouter/<provider>/<model>
    ```

  </Step>
</Steps>

Config example

{
  env: { vars: { OPENROUTER_API_KEY: "sk-or-..." } },
  agents: {
    defaults: {
      model: { primary: "openrouter/auto" },
    },
  },
}

Model references

Model refs follow the pattern `openrouter//`. For the full list of providers and models OpenRouter routes to, see [OpenRouter's model catalog](https://openrouter.ai/models). For how OpenClaw resolves model refs and failover, see [Model selection](/concepts/model-providers).

Bundled starter models enrich a nonempty public catalog. A failed live request reports a discovery failure rather than substituting these rows; a successful empty response stays empty:

Model ref	Notes
openrouter/auto	OpenRouter automatic routing
openrouter/moonshotai/kimi-k2.6	Kimi K2.6 via MoonshotAI
openrouter/moonshotai/kimi-k2.5	Kimi K2.5 via MoonshotAI

Any other openrouter/<provider>/<model> ref, including openrouter/openrouter/fusion (see Fusion router), resolves dynamically against OpenRouter's live model catalog.

Discovered models use OpenRouter's advertised tool support. When a model's supported_parameters list omits tools, OpenClaw sends requests without tool definitions or tool choice. Models without that metadata keep the default tool behavior.

Image generation

OpenRouter can back the image_generate tool. Set an OpenRouter image model under agents.defaults.mediaModels.image:

{
  env: { vars: { OPENROUTER_API_KEY: "sk-or-..." } },
  agents: {
    defaults: {
      mediaModels: {
        image: {
          primary: "openrouter/google/gemini-3.1-flash-image-preview",
          timeoutMs: 180000,
        },
      },
    },
  },
}

OpenClaw sends standard OpenRouter image requests to the dedicated image API (POST /api/v1/images). Gemini image models additionally receive aspect_ratio and resolution hints, and image edits pass source images as input_references. Generated images come back as base64 (b64_json) with an optional media_type; when media_type is absent, OpenClaw sniffs the image format from the bytes.

Configured custom OpenRouter baseUrl destinations retain the existing chat-completions image route for compatibility with proxies that do not expose the dedicated endpoint. Use agents.defaults.mediaModels.image.timeoutMs for slower models; the image_generate tool's per-call timeoutMs still wins.

Video generation

OpenRouter can back the video_generate tool through its asynchronous /videos API. Set an OpenRouter video model under agents.defaults.mediaModels.video:

{
  env: { vars: { OPENROUTER_API_KEY: "sk-or-..." } },
  agents: {
    defaults: {
      mediaModels: {
        video: {
          primary: "openrouter/google/veo-3.1-fast",
        },
      },
    },
  },
}

OpenClaw submits text-to-video and image-to-video jobs, polls the returned polling_url, and downloads the finished video from OpenRouter's unsigned_urls or the job content endpoint. Reference images default to first/last-frame images; images tagged reference_image are sent as input references instead. The bundled google/veo-3.1-fast default supports 4/6/8 second durations, 720P/1080P resolutions, and 16:9/9:16 aspect ratios. Video-to-video is not supported: the upstream API only accepts text and image references.

Music generation

OpenRouter can back the music_generate tool through chat-completions audio output. Set an OpenRouter audio model under agents.defaults.mediaModels.music:

{
  env: { vars: { OPENROUTER_API_KEY: "sk-or-..." } },
  agents: {
    defaults: {
      mediaModels: {
        music: {
          primary: "openrouter/google/lyria-3-pro-preview",
          timeoutMs: 180000,
        },
      },
    },
  },
}

The bundled OpenRouter music provider defaults to google/lyria-3-pro-preview and also exposes google/lyria-3-clip-preview. OpenClaw sends modalities: ["text", "audio"], streams the response, collects the audio chunks, and saves the result as generated media for channel delivery. Lyria models accept one reference image through the shared music_generate image=... parameter. Streaming audio, transcript retention, and the derived SSE event envelope are bounded by agents.defaults.mediaMaxMb (the default audio cap is 16 MB).

Text-to-speech

OpenRouter can act as a TTS provider through its OpenAI-compatible /audio/speech endpoint.

{
  tts: {
    auto: "always",
    provider: "openrouter",
    providers: {
      openrouter: {
        model: "hexgrad/kokoro-82m",
        speakerVoice: "af_alloy",
        responseFormat: "mp3",
      },
    },
  },
}

If tts.providers.openrouter.apiKey is omitted, TTS falls back to models.providers.openrouter.apiKey, then OPENROUTER_API_KEY.

Speech-to-text (inbound audio)

OpenRouter can transcribe inbound voice/audio attachments through the shared tools.media.audio path, using its STT endpoint (/audio/transcriptions). This applies to any channel plugin that forwards inbound voice/audio into media understanding preflight.

{
  tools: {
    media: {
      models: [
        {
          provider: "openrouter",
          model: "openai/whisper-large-v3-turbo",
          capabilities: ["audio"],
        },
      ],
      audio: { enabled: true },
    },
  },
}

OpenClaw sends OpenRouter STT requests as JSON with base64 audio under input_audio (OpenRouter's STT contract), not as multipart OpenAI form uploads.

Fusion router

OpenRouter Fusion sends one OpenClaw model ref to several OpenRouter models in parallel, has OpenRouter judge their answers, and returns one final response through the normal OpenRouter endpoint. The upstream model slug is openrouter/fusion, so the OpenClaw model ref carries both the OpenClaw provider prefix and the upstream OpenRouter namespace:

openclaw models set openrouter/openrouter/fusion

Configure Fusion's panel and judge through the model's params.extraBody; those fields forward directly into the OpenRouter chat-completions request body. Fusion works with either OAuth or API-key onboarding; if you use OAuth, omit the env.vars.OPENROUTER_API_KEY line below.

{
  env: { vars: { OPENROUTER_API_KEY: "sk-or-..." } },
  agents: {
    defaults: {
      model: { primary: "openrouter/openrouter/fusion" },
      models: {
        "openrouter/openrouter/fusion": {
          params: {
            extraBody: {
              plugins: [
                {
                  id: "fusion",
                  analysis_models: [
                    "google/gemini-3.5-flash",
                    "moonshotai/kimi-k2.6",
                    "deepseek/deepseek-v4-pro",
                  ],
                  model: "google/gemini-3.5-flash",
                },
              ],
            },
          },
        },
      },
    },
  },
}

analysis_models is the parallel panel; model inside the Fusion plugin config is the judge model. Do not set top-level tool_choice to "required" in normal agent/chat turns to try to force Fusion: OpenClaw turns can include its own tool definitions, and a top-level required tool choice may pick one of those instead of the Fusion router. When this Fusion plugin config is present, OpenClaw adds a sanitized system-prompt note listing the configured analysis models and judge model, so the agent can answer questions about its own Fusion panel. Other extraBody fields are not copied into the prompt.

Fusion is slower by design: OpenRouter fans the prompt out to multiple analysis models, then runs a judge/synthesis step, so latency runs higher than a direct single-model request. Use it for deliberate, high-quality answers or escalation paths, not as a latency-sensitive default. Keep the panel small and pick faster analysis/judge models for quicker responses.

Test a configured ref with a one-shot local call:

openclaw infer model run --local \
  --model openrouter/openrouter/fusion \
  --prompt "Reply with exactly: FUSION_OK" \
  --json

Authentication and headers

OpenRouter uses a Bearer token from your API key. OpenRouter OAuth is a PKCE login flow that issues an OpenRouter API key, so OpenClaw stores the result in the same openrouter:default API-key auth profile used by manual API-key setup.

To sign in or rotate the stored key on an existing install without rerunning full onboarding:

openclaw models auth login --provider openrouter --method oauth
openclaw models auth login --provider openrouter --method api-key

On requests to OpenRouter endpoints (openrouter.ai), OpenClaw adds OpenRouter's documented app-attribution headers. This applies to the bundled openrouter provider and to custom provider ids whose baseUrl points at OpenRouter:

Header	Value
HTTP-Referer	https://openclaw.ai
X-OpenRouter-Title	OpenClaw
X-OpenRouter-Categories	personal-agent,cli-agent

If you repoint the OpenRouter provider at some other proxy or base URL, OpenClaw does **not** inject those OpenRouter-specific headers or Anthropic cache markers.

Advanced configuration

OpenRouter response caching is opt-in. Enable it per model:

```json5
{
  agents: {
    defaults: {
      models: {
        "openrouter/auto": {
          params: {
            responseCache: true,
            responseCacheTtlSeconds: 300,
          },
        },
      },
    },
  },
}
```

OpenClaw sends `X-OpenRouter-Cache: true` and, when configured,
`X-OpenRouter-Cache-TTL`. `responseCacheClear: true` forces a refresh for
the current request and stores the replacement response. Snake_case
aliases (`response_cache`, `response_cache_ttl_seconds`,
`response_cache_clear`) are accepted, as is `responseCacheTtl` /
`response_cache_ttl` without the `Seconds` suffix.
TTL values are truncated and clamped to 1–86400 seconds; non-finite values are ignored.

This is separate from provider prompt caching and from OpenRouter's
Anthropic `cache_control` markers. It only applies on verified
`openrouter.ai` routes, not custom proxy base URLs.

On verified OpenRouter routes, Anthropic model refs keep OpenRouter's Anthropic `cache_control` markers for better prompt-cache reuse on system/developer prompt blocks. On verified OpenRouter routes, Anthropic model refs with reasoning enabled drop trailing assistant prefill turns before the request reaches OpenRouter, matching Anthropic's requirement that reasoning conversations end with a user turn. OpenClaw uses the selected model's advertised reasoning efforts for its thinking choices and request payloads. Models that require reasoning omit the off choice. Agent turns and standalone completions share these controls and reasoning-replay rules. On supported non-`auto` routes, OpenClaw maps the selected thinking level to OpenRouter proxy reasoning payloads. `openrouter/auto` and unsupported model hints skip that injection. Stale `openrouter/hunter-alpha` refs also skip it, because OpenRouter could return final answer text in reasoning fields on that retired route.

Models without an effort selector show on/off controls, or **always on**
when reasoning is mandatory. These models receive binary reasoning controls
without a scalar effort. Omitting a thinking request leaves their native
reasoning default unchanged; configured reasoning budgets are preserved.

On verified OpenRouter routes, `openrouter/deepseek/deepseek-v4-flash` and `openrouter/deepseek/deepseek-v4-pro` fill missing `reasoning_content` on replayed assistant turns, keeping thinking/tool conversations in DeepSeek V4's required follow-up shape. OpenClaw sends OpenRouter-supported `reasoning.effort` values for these routes: `xhigh`/`max` map to `xhigh`, every other non-off level maps to `high`. `/think off` explicitly sends `reasoning.effort: "none"` and removes reasoning replay fields instead of falling back to the provider's reasoning default. OpenRouter runs through the proxy-style OpenAI-compatible path, so native OpenAI-only request shaping such as `serviceTier`, Responses `store`, OpenAI reasoning-compat payloads, and prompt-cache hints is not forwarded. Gemini-backed OpenRouter refs stay on the proxy-Gemini path: OpenClaw keeps Gemini thought-signature sanitation there, but does not enable native Gemini replay validation or bootstrap rewrites. OpenRouter supports a `provider` request object for underlying provider routing. Configure a default policy for all OpenRouter text-model requests with `models.providers.openrouter.params.provider`:

```json5
{
  models: {
    providers: {
      openrouter: {
        params: {
          provider: {
            sort: "latency",
            require_parameters: true,
            data_collection: "deny",
          },
        },
      },
    },
  },
}
```

OpenClaw forwards that object to OpenRouter as the request `provider`
payload. Use OpenRouter's documented snake_case fields, including `sort`,
`only`, `ignore`, `order`, `allow_fallbacks`, `require_parameters`,
`data_collection`, `quantizations`, `max_price`, `preferred_max_latency`,
`preferred_min_throughput`, `zdr`, and `enforce_distillable_text`.

Per-model params override the provider-wide routing object:

```json5
{
  agents: {
    defaults: {
      models: {
        "openrouter/anthropic/claude-sonnet-4-6": {
          params: {
            provider: {
              order: ["anthropic"],
              allow_fallbacks: false,
            },
          },
        },
      },
    },
  },
}
```

This only applies on OpenRouter chat-completions routes. Direct Anthropic,
Google, OpenAI, or custom provider routes ignore OpenRouter routing params.

Related

Choosing providers, model refs, and failover behavior. Full config reference for agents, models, and providers. Arcee models reachable with an OpenRouter key. Shared image tool parameters and provider selection. Shared video tool parameters and provider selection. Shared music tool parameters and provider selection.