screenscribe-mcp

# screenscribe **Extract *typed* data from videos — and synthesize across many.** The best how-to knowledge — live demos, conference talks, tutorials, cooking videos — is trapped in video: you can watch it, but you can't *use* it. screenscribe turns it into structured, validated data an agent can act on. Hand it a schema and it fills it from what's on screen *and* said; point it at a whole channel and it synthesizes one coherent artifact across every video. **Powered by Gemini (it watches the video); your agent does the reasoning.** Standalone CLI + MCP server (Claude Code, Cursor, Windsurf, any MCP client). One key — and transcript-only mode needs none. Works regardless of the video's language and needs no captions. --- ## What it does — three layers 1. **Watch & extract.** Gemini watches the video. Pull the transcript (free), a cheap whole-video analysis, or the key frames as PNG images your agent opens and reads. 2. **Typed extraction** ⭐ — hand it a **JSON Schema** (or a bundled preset) and get back **validated JSON**: every CLI command shown, the final config file, a recipe with quantities, the code at each step. Prose is for humans to read; this is what an agent automates on. 3. **Cross-video synthesis** ⭐ — point at a **channel / playlist / list of URLs** → fan the typed extraction out over every video → **compound** the results into one artifact conforming to an *aggregate* schema (a cookbook, a technique grammar, a comparison). It scales past a single context window, one capped batch at a time. The raw extraction is a commodity (it's Gemini under the hood). The durable value is the **structure** and the **cross-video synthesis** — and both only get better as the models do. --- ## Why it matters - **Watch → automate.** *"Extract every command shown in this setup tutorial"* → a typed list your tooling runs. *"Turn mom's whole cooking channel into one cookbook"* → a synthesized artifact, not 300 summaries. - **Typed, not prose.** Output validates ag…