Engineer by craft. Builder by instinct.Engineer by craft. Builder by instinct.ctxr
Turn YouTube videos into agent-readable context: transcripts, keyframes, and timestamped Markdown walkthroughs, available through a CLI and MCP server.
Give agents the context inside a video.
Useful product knowledge often lives in demos, tutorials, and talks. A transcript alone loses the interface being shown; isolated screenshots lose the explanation. ctxr aligns the two into material an agent can read.
04 / Technologies & context
- Python
- AI Agents
- MCP
- Video
- Open Source
Inside the project
Implementation
The public README describes a Python pipeline using yt-dlp for downloads, ffmpeg for scene-based frame extraction, and captions with optional local Whisper transcription. Transcript segments are aligned to frames and written as Markdown and structured manifests. An MCP server exposes processing and focused queries over the output.
The workflow
- Choose a YouTube video, playlist, or page of embedded videos
- Extract captions or a local transcript and representative keyframes
- Read or query timestamped walkthroughs through the CLI and MCP server
Scope & status
Repository documentation reviewed, not independently runtime-tested. YouTube is the documented source platform. Automatic transcripts can mishear names, and source availability and rate limits affect processing. No accuracy or throughput guarantee is claimed.
Public references
Let's build
something
together.
Have a problem worth solving? I want to hear about it.