<p align="center">
  <img src="./assets/hero.png" alt="See What I Mean — give text-only agents eyes and ears">
</p>

# See What I Mean

Give text-only Claude Code and Codex agents image, video, and audio analysis through a local MCP server. Choose free GLM-4.6V-Flash for images and keyframe video, or MiMo V2.5 for images, direct video, and audio.

The user instruction stays one line:

```text
查看并安装这个 Skill：https://github.com/ov-c/see-what-i-mean
```

The agent owns installation, provider selection, API-key collection, configuration, verification, and any required restart. The user should never need to copy commands or edit paths.

## Codex agent procedure

1. Ensure Python 3.10+ and `ffmpeg` are available.
2. Run `codex plugin marketplace add ov-c/see-what-i-mean`, then `codex plugin add see-what-i-mean@see-what-i-mean`.
3. Do not also copy the standalone Skill or run `codex mcp add`; the plugin already owns both discovery and MCP registration.
4. In a new task, follow the loaded `see-what-i-mean` Skill. Ask only for provider choice and the matching API key.

The Codex plugin uses a relative MCP launcher and creates a versioned private Python runtime on first start. It does not publish a machine-specific Python path. Verify native tool discovery from the Codex tool inventory; an inline Python MCP client is not an acceptable substitute.

## Claude Code agent procedure

1. Add this repository as a plugin marketplace and install `see-what-i-mean@see-what-i-mean-marketplace`.
2. Follow the loaded Skill and run `python scripts/agent_bootstrap.py --client claude --registration plugin` from the plugin root.
3. Request a restart only if the four MCP tools are still unavailable.

## Providers and tools

- GLM-4.6V-Flash: `analyze_image` and eight-keyframe `analyze_video`; no documented audio input.
- MiMo V2.5: `analyze_image`, direct video up to 30 MB, keyframe fallback for larger video, and `analyze_audio`.
- `delegate_text`: optional bounded text preprocessing through a user-configured non-interactive CLI such as `claude -p`.

Media is uploaded to the selected provider. API keys are stored only in the user's private configuration. Never commit keys or share unredacted conversation exports.