talkthrough-mcp: Local MCP Server Turns Screen Recordings into AI Data

talkthrough-mcp is an MIT-licensed, local-first MCP server that converts narrated screen recordings—or any video/audio file—into structured data agents can use: timestamped transcript segments, optional speaker labels, scene-change keyframes, OCR'd on-screen text, and wall-clock anchoring. There's no LLM or cloud dependency inside the server; ffmpeg, faster-whisper, and RapidOCR run entirely on-device, while the calling agent supplies the reasoning.

The server exposes lazy retrieval tools so a 30-minute recording never floods an agent's context window—only the exact transcript slice, moment bundle, or frame needed gets pulled. It installs via a single uvx command across MCP clients including Claude Code, Cursor, Codex CLI, Gemini CLI, Cline, and Zed, with an accompanying plugin offering slash commands and a triage subagent.

Bundled server prompts target concrete workflows: turning a narrated screencast into a findings report, extracting a spec from a recorded workshop, building a backlog from a product demo, pulling action items from meeting audio, and correlating spoken remarks with system logs via wall-clock timestamps—useful for QA, product, and support teams processing raw recordings at scale.