How to add vision/image analysis and YouTube support to OpenClaw?

[delicate token]

Hi, I'm running OpenClaw on WSL2 with local Ollama models.

Two things I want to add:

1. Image/screenshot analysis - my agent can't currently analyze images
2. YouTube/video content analysis

I have qwen3-vl:4b-instruct running locally which should support vision.

What's the recommended way to enable these capabilities? Any skills or config changes needed?

[deft flame]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[violet lintel]

Let me check the current docs and source so I don’t make up config keys.

Short version, image analysis yes, YouTube/video is a separate thing.

1. Image / screenshot analysis
   • You do not need a skill for this.
   • Recommended path is to use the native Ollama provider, not Ollama’s /v1 OpenAI-compatible URL.
   • Then either:
     • let OpenClaw auto-detect the vision model, or
     • set it explicitly as the image model:

openclaw models list --provider ollama
openclaw models set-image ollama/qwen3-vl:4b-instruct
openclaw models status

Important gotcha:

• If you configured models.providers.ollama manually, auto-discovery is skipped.

• In that case your model entry must be marked as image-capable, otherwise OpenClaw will treat it as text-only.
• Also use baseUrl: "http://...:11434" with api: "ollama", not /v1.

2. Video / YouTube analysis
   • Uploaded video clips: built-in media understanding can handle these, but current provider support is for Google, Qwen (DashScope), and Moonshot. This is not the same as local Ollama Qwen3-VL.
   • YouTube URLs specifically: there is no turnkey built-in “paste YouTube link and summarize” flow in core OpenClaw right now.

So the practical recommendation is:

• For screenshots/images: use your local Ollama qwen3-vl:4b-instruct via agents.defaults.imageModel.
• For video files: use a supported remote video-understanding provider.
• For YouTube links: add a custom skill/plugin that fetches captions or downloads/transcribes the video, then feeds that text/video into the agent.

Docs:

• Ollama: https://docs.openclaw.ai/providers/ollama
• Media understanding: https://docs.openclaw.ai/nodes/media-understanding
• Models / imageModel: https://docs.openclaw.ai/concepts/models

If you want, paste:

openclaw --version
openclaw models status

and I can tell you exactly what to change in your config.

[delicate token]

My primary model is myclaw/minimax-m2.7. Does it support image/vision input? If not, how do I configure a separate image model for vision tasks?

[violet lintel]

Probably not by default, at least not in the way OpenClaw routes vision.

Two cases:

1. If myclaw/minimax-m2.7 is the bundled MiniMax provider

   • M2.7 is generally treated as the text chat model.
   • MiniMax image understanding is exposed separately via MiniMax-VL-01.
   • So I would not rely on M2.7 as your vision model.

2. If myclaw/minimax-m2.7 is your own custom provider/model entry

   • OpenClaw only treats it as vision-capable if that model is declared with:

   input: ["text", "image"]
   

   • Otherwise it is treated as text-only.

Best setup for your case

Keep your main text model as-is, and set a separate image model:

openclaw models set-image ollama/qwen3-vl:4b-instruct

Or in config:

{
  agents: {
    defaults: {
      model: {
        primary: "myclaw/minimax-m2.7",
      },
      imageModel: {

        primary: "ollama/qwen3-vl:4b-instruct",
      },
    },
  },
}

That is exactly what imageModel is for. OpenClaw uses it when the active/default model cannot accept image input.

Important Ollama note

For the Ollama side, use the native Ollama API, not /v1:

{

  models: {
    providers: {
      ollama: {
        baseUrl: "http://127.0.0.1:11434",
        api: "ollama",
      },
    },
  },
}

If qwen3-vl:4b-instruct is auto-discovered from Ollama, OpenClaw should mark it image-capable automatically.

Docs:

• https://docs.openclaw.ai/concepts/models
• https://docs.openclaw.ai/providers/ollama