Media understanding

OpenClaw can summarize inbound media (image/audio/video) before the reply pipeline runs. It auto-detects when local tools or provider keys are available, and can be disabled or customized. If understanding is off, models still receive the original files/URLs as usual.

Vendor-specific media behavior is registered by vendor plugins, while OpenClaw core owns the shared tools.media config, fallback order, and reply-pipeline integration.

Goals

Optional: pre-digest inbound media into short text for faster routing + better command parsing.
Preserve original media delivery to the model (always).
Support provider APIs and CLI fallbacks.
Allow multiple models with ordered fallback (error/size/timeout).

High-level behavior

<Steps> <Step title="Collect attachments"> Collect inbound attachments (`MediaPaths`, `MediaUrls`, `MediaTypes`). </Step> <Step title="Select per-capability"> For each enabled capability (image/audio/video), select attachments per policy (default: **first**). </Step> <Step title="Choose model"> Choose the first eligible model entry (size + capability + auth). </Step> <Step title="Fallback on failure"> If a model fails or the media is too large, **fall back to the next entry**. </Step> <Step title="Apply success block"> On success:

- `Body` becomes `[Image]`, `[Audio]`, or `[Video]` block.
- Audio sets `{{Transcript}}`; command parsing uses caption text when present, otherwise the transcript.
- Captions are preserved as `User text:` inside the block.

</Step> </Steps>

If understanding fails or is disabled, the reply flow continues with the original body + attachments.

Config overview

tools.media supports shared models plus per-capability overrides:

<AccordionGroup> <Accordion title="Top-level keys"> - `tools.media.models`: shared model list (use `capabilities` to gate). - `tools.media.image` / `tools.media.audio` / `tools.media.video`: - defaults (`prompt`, `maxChars`, `maxBytes`, `timeoutSeconds`, `language`) - provider overrides (`baseUrl`, `headers`, `providerOptions`) - Deepgram audio options via `tools.media.audio.providerOptions.deepgram` - audio transcript echo controls (`echoTranscript`, default `false`; `echoFormat`) - optional **per-capability `models` list** (preferred before shared models) - `attachments` policy (`mode`, `maxAttachments`, `prefer`) - `scope` (optional gating by channel/chatType/session key) - `tools.media.concurrency`: max concurrent capability runs (default **2**). </Accordion> </AccordionGroup>

json5

{
  tools: {
    media: {
      models: [
        /* shared list */
      ],
      image: {
        /* optional overrides */
      },
      audio: {
        /* optional overrides */
        echoTranscript: true,
        echoFormat: '📝 "{transcript}"',
      },
      video: {
        /* optional overrides */
      },
    },
  },
}

Model entries

Each models[] entry can be provider or CLI:

<Tabs> <Tab title="Provider entry"> ```json5 { type: "provider", // default if omitted provider: "openai", model: "gpt-5.5", prompt: "Describe the image in <= 500 chars.", maxChars: 500, maxBytes: 10485760, timeoutSeconds: 60, capabilities: ["image"], // optional, used for multi-modal entries profile: "vision-profile", preferredProfile: "vision-fallback", } ``` </Tab> <Tab title="CLI entry"> ```json5 { type: "cli", command: "gemini", args: [ "-m", "gemini-3-flash", "--allowed-tools", "read_file", "Read the media at {{MediaPath}} and describe it in <= {{MaxChars}} characters.", ], maxChars: 500, maxBytes: 52428800, timeoutSeconds: 120, capabilities: ["video", "image"], } ```

CLI templates can also use:

- `{{MediaDir}}` (directory containing the media file)
- `{{OutputDir}}` (scratch dir created for this run)
- `{{OutputBase}}` (scratch file base path, no extension)

</Tab> </Tabs>

Defaults and limits

Recommended defaults:

maxChars: 500 for image/video (short, command-friendly)
maxChars: unset for audio (full transcript unless you set a limit)
maxBytes:
- image: 10MB
- audio: 20MB
- video: 50MB

<AccordionGroup> <Accordion title="Rules"> - If media exceeds `maxBytes`, that model is skipped and the **next model is tried**. - Audio files smaller than **1024 bytes** are treated as empty/corrupt and skipped before provider/CLI transcription; inbound reply context receives a deterministic placeholder transcript so the agent knows the note was too small. - If the model returns more than `maxChars`, output is trimmed. - `prompt` defaults to simple "Describe the {media}." plus the `maxChars` guidance (image/video only). - If the active primary image model already supports vision natively, OpenClaw skips the `[Image]` summary block and passes the original image into the model instead. - If a Gateway/WebChat primary model is text-only, image attachments are preserved as offloaded `media://inbound/*` refs so the image/PDF tools or configured image model can still inspect them instead of losing the attachment. - Explicit `openclaw infer image describe --model <provider/model>` requests are different: they run that image-capable provider/model directly, including Ollama refs such as `ollama/qwen2.5vl:7b`. - If `<capability>.enabled: true` but no models are configured, OpenClaw tries the **active reply model** when its provider supports the capability. </Accordion> </AccordionGroup>

Auto-detect media understanding (default)

If tools.media.<capability>.enabled is not set to false and you haven't configured models, OpenClaw auto-detects in this order and stops at the first working option:

<Steps> <Step title="Active reply model"> Active reply model when its provider supports the capability. </Step> <Step title="agents.defaults.imageModel"> `agents.defaults.imageModel` primary/fallback refs (image only). Prefer `provider/model` refs. Bare refs are qualified from configured image-capable provider model entries only when the match is unique. </Step> <Step title="Local CLIs (audio only)"> Local CLIs (if installed):

- `sherpa-onnx-offline` (requires `SHERPA_ONNX_MODEL_DIR` with encoder/decoder/joiner/tokens)
- `whisper-cli` (`whisper-cpp`; uses `WHISPER_CPP_MODEL` or the bundled tiny model)
- `whisper` (Python CLI; downloads models automatically)

</Step> <Step title="Gemini CLI"> `gemini` using `read_many_files`. </Step> <Step title="Provider auth"> - Configured `models.providers.*` entries that support the capability are tried before the bundled fallback order. - Image-only config providers with an image-capable model auto-register for media understanding even when they are not a bundled vendor plugin. - Ollama image understanding is available when selected explicitly, for example through `agents.defaults.imageModel` or `openclaw infer image describe --model ollama/<vision-model>`.

Bundled fallback order:

- Audio: OpenAI → Groq → xAI → Deepgram → Google → SenseAudio → ElevenLabs → Mistral
- Image: OpenAI → Anthropic → Google → MiniMax → MiniMax Portal → Z.AI
- Video: Google → Qwen → Moonshot

</Step> </Steps>

To disable auto-detection, set:

json5

{
  tools: {
    media: {
      audio: {
        enabled: false,
      },
    },
  },
}

<Note> Binary detection is best-effort across macOS/Linux/Windows; ensure the CLI is on `PATH` (we expand `~`), or set an explicit CLI model with a full command path. </Note>

Proxy environment support (provider models)

When provider-based audio and video media understanding is enabled, OpenClaw honors standard outbound proxy environment variables for provider HTTP calls:

HTTPS_PROXY
HTTP_PROXY
ALL_PROXY
https_proxy
http_proxy
all_proxy

If no proxy env vars are set, media understanding uses direct egress. If the proxy value is malformed, OpenClaw logs a warning and falls back to direct fetch.

Capabilities (optional)

If you set capabilities, the entry only runs for those media types. For shared lists, OpenClaw can infer defaults:

openai, anthropic, minimax: image
minimax-portal: image
moonshot: image + video
openrouter: image
google (Gemini API): image + audio + video
qwen: image + video
mistral: audio
zai: image
groq: audio
xai: audio
deepgram: audio
Any models.providers.<id>.models[] catalog with an image-capable model: image

For CLI entries, set capabilities explicitly to avoid surprising matches. If you omit capabilities, the entry is eligible for the list it appears in.

Provider support matrix (OpenClaw integrations)

Capability	Provider integration	Notes
Image	OpenAI, OpenAI Codex OAuth, Codex app-server, OpenRouter, Anthropic, Google, MiniMax, Moonshot, Qwen, Z.AI, config providers	Vendor plugins register image support; `openai-codex/` uses OAuth provider plumbing; `codex/` uses a bounded Codex app-server turn; MiniMax and MiniMax OAuth both use `MiniMax-VL-01`; image-capable config providers auto-register.
Audio	OpenAI, Groq, xAI, Deepgram, Google, SenseAudio, ElevenLabs, Mistral	Provider transcription (Whisper/Groq/xAI/Deepgram/Gemini/SenseAudio/Scribe/Voxtral).
Video	Google, Qwen, Moonshot	Provider video understanding via vendor plugins; Qwen video understanding uses the Standard DashScope endpoints.

<Note> **MiniMax note**

minimax and minimax-portal image understanding comes from the plugin-owned MiniMax-VL-01 media provider.
The bundled MiniMax text catalog still starts text-only; explicit models.providers.minimax entries materialize image-capable M2.7 chat refs.

</Note>

Model selection guidance

Prefer the strongest latest-generation model available for each media capability when quality and safety matter.
For tool-enabled agents handling untrusted inputs, avoid older/weaker media models.
Keep at least one fallback per capability for availability (quality model + faster/cheaper model).
CLI fallbacks (whisper-cli, whisper, gemini) are useful when provider APIs are unavailable.
parakeet-mlx note: with --output-dir, OpenClaw reads <output-dir>/<media-basename>.txt when output format is txt (or unspecified); non-txt formats fall back to stdout.

Attachment policy

Per-capability attachments controls which attachments are processed:

<ParamField path="mode" type='"first" | "all"' default="first"> Whether to process the first selected attachment or all of them. </ParamField> <ParamField path="maxAttachments" type="number" default="1"> Cap the number processed. </ParamField> <ParamField path="prefer" type='"first" | "last" | "path" | "url"'> Selection preference among candidate attachments. </ParamField>

When mode: "all", outputs are labeled [Image 1/2], [Audio 2/2], etc.

<AccordionGroup> <Accordion title="File-attachment extraction behavior"> - Extracted file text is wrapped as **untrusted external content** before it is appended to the media prompt. - The injected block uses explicit boundary markers like `<<<EXTERNAL_UNTRUSTED_CONTENT id="...">>>` / `<<<END_EXTERNAL_UNTRUSTED_CONTENT id="...">>>` and includes a `Source: External` metadata line. - This attachment-extraction path intentionally omits the long `SECURITY NOTICE:` banner to avoid bloating the media prompt; the boundary markers and metadata still remain. - If a file has no extractable text, OpenClaw injects `[No extractable text]`. - If a PDF falls back to rendered page images in this path, the media prompt keeps the placeholder `[PDF content rendered to images; images not forwarded to model]` because this attachment-extraction step forwards text blocks, not the rendered PDF images. </Accordion> </AccordionGroup>

Config examples

<Tabs> <Tab title="Shared models + overrides"> ```json5 { tools: { media: { models: [ { provider: "openai", model: "gpt-5.5", capabilities: ["image"] }, { provider: "google", model: "gemini-3-flash-preview", capabilities: ["image", "audio", "video"], }, { type: "cli", command: "gemini", args: [ "-m", "gemini-3-flash", "--allowed-tools", "read_file", "Read the media at {{MediaPath}} and describe it in <= {{MaxChars}} characters.", ], capabilities: ["image", "video"], }, ], audio: { attachments: { mode: "all", maxAttachments: 2 }, }, video: { maxChars: 500, }, }, }, } ``` </Tab> <Tab title="Audio + video only"> ```json5 { tools: { media: { audio: { enabled: true, models: [ { provider: "openai", model: "gpt-4o-mini-transcribe" }, { type: "cli", command: "whisper", args: ["--model", "base", "{{MediaPath}}"], }, ], }, video: { enabled: true, maxChars: 500, models: [ { provider: "google", model: "gemini-3-flash-preview" }, { type: "cli", command: "gemini", args: [ "-m", "gemini-3-flash", "--allowed-tools", "read_file", "Read the media at {{MediaPath}} and describe it in <= {{MaxChars}} characters.", ], }, ], }, }, }, } ``` </Tab> <Tab title="Image-only"> ```json5 { tools: { media: { image: { enabled: true, maxBytes: 10485760, maxChars: 500, models: [ { provider: "openai", model: "gpt-5.5" }, { provider: "anthropic", model: "claude-opus-4-6" }, { type: "cli", command: "gemini", args: [ "-m", "gemini-3-flash", "--allowed-tools", "read_file", "Read the media at {{MediaPath}} and describe it in <= {{MaxChars}} characters.", ], }, ], }, }, }, } ``` </Tab> <Tab title="Multi-modal single entry"> ```json5 { tools: { media: { image: { models: [ { provider: "google", model: "gemini-3.1-pro-preview", capabilities: ["image", "video", "audio"], }, ], }, audio: { models: [ { provider: "google", model: "gemini-3.1-pro-preview", capabilities: ["image", "video", "audio"], }, ], }, video: { models: [ { provider: "google", model: "gemini-3.1-pro-preview", capabilities: ["image", "video", "audio"], }, ], }, }, }, } ``` </Tab> </Tabs>

Status output

When media understanding runs, /status includes a short summary line:

📎 Media: image ok (openai/gpt-5.4) · audio skipped (maxBytes)

This shows per-capability outcomes and the chosen provider/model when applicable.

Notes

Understanding is best-effort. Errors do not block replies.
Attachments are still passed to models even when understanding is disabled.
Use scope to limit where understanding runs (e.g. only DMs).