watch-skill — Claude Code MCP server: what it does, how to install
Indexes video, audio, and screen sessions into timestamped records. Gives AI agents deterministic contracts to verify task completion.
claude mcp add watch-skill -- uvx watch-skillWhat watch-skill does
The server parses screen activity, audio tracks, and video clips into searchable frames, text transcripts, and OCR data. Each captured observation attaches to an exact timestamp. When an agent queries indexed footage, it returns evidence referencing a verifiable second in the recording.
Beyond observation, the engine verifies test contracts without trusting model self-evaluations. An independent runtime evaluates DOM snapshots, JSON payloads, file digests, HTTP responses, and SQL queries. The evaluation finishes with one of four outcomes: VERIFIED, FAILED, UNVERIFIED, or INCONCLUSIVE. It exposes 39 tools over stdio.
A minimal pip install watch-skill cannot grab video frames on its own. It throws an error unless installed with extra dependencies like watch-skill[standard] or watch-skill[all]. Running watch-skill doctor detects missing system binaries and fixes what it can.
What it helps with
- Locate specific UI transitions in screen recordings using exact timestamps
- Verify API JSON responses and file digests against deterministic criteria
- Extract on-screen text from video sessions using OCR
- Validate browser DOM states without relying on language model opinions
Example prompts for the agent
- Inspect screen.mp4 and check what broke during user registration
- Verify the database contract for the migration task
- Identify what happened on screen at 03:12 in demo.webm
How to install watch-skill
Platform CLI
claude mcp add watch-skill -- uvx watch-skilluvx runs the command named after the package; if the README names another one, use uvx --from watch-skill <command>
Platform CLI
codex mcp add watch-skill -- uvx watch-skilluvx runs the command named after the package; if the README names another one, use uvx --from watch-skill <command>
Config file
{
"mcpServers": {
"watch-skill": {
"command": "uvx",
"args": [
"watch-skill"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from watch-skill <command>
Config file
{
"servers": {
"watch-skill": {
"type": "stdio",
"command": "uvx",
"args": [
"watch-skill"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from watch-skill <command>
Platform CLI
gemini mcp add watch-skill uvx watch-skilluvx runs the command named after the package; if the README names another one, use uvx --from watch-skill <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about watch-skill
How do I install watch-skill in Claude Code?
Run claude mcp add watch-skill -- uvx watch-skill in your terminal. You will need uv installed in your environment.
What does watch-skill do?
It supplies agents with 39 tools to process multimedia records and screen sessions. It also evaluates deterministic contracts like file hashes and DOM state to prove whether a task succeeded.
Why does video ingestion fail after a bare pip install?
A base installation omits frame processing dependencies and halts on video analysis. You need optional packages such as watch-skill[standard] or watch-skill[all], and you can verify requirements using watch-skill doctor.
