glideit

Your coding agent can now watch long videos.

Paste a YouTube link (or any video URL, or a local file) and ask a question. glideit downloads the video, builds a transcript and a storyboard of the whole thing, then extracts high-res frames of just the part that matters. Your agent reads them and answers — grounded in what is actually on screen.

▶ Watch the full launch video — with narration and score (38s, made with HyperFrames from demo/demo.html)

• No API keys. No cloud. Everything runs locally: ffmpeg, yt-dlp, optional tesseract. The agent that invoked glideit does all the "seeing" — no model API is ever called.

• Built for long videos. A 1-hour lecture becomes 4 storyboard images + a transcript, not 450 frames flooding the agent's context.

• Reads on-screen code. High-res zoom frames + an OCR sidecar make IDE/terminal content legible.

Install

Claude Code:

/plugin marketplace add Imhari14/glideit
/plugin install glideit@glideit

When the installer asks for a scope, pick User (the default) so /glideit works in every project — Project scope pins it to the folder that's open. Non-interactive: claude plugin install glideit@glideit --scope user. Installed with Project scope by mistake? Reinstall: claude plugin uninstall glideit@glideit --scope project, then claude plugin install glideit@glideit --scope user. New installs load on the next session — restart or reload the window.

Cursor, Codex, Copilot, Gemini CLI, and 70+ other agents:

npx skills add Imhari14/glideit -g

Requirements: Python 3.10+, ffmpeg, yt-dlp (pip install yt-dlp). Run python scripts/setup.py to check. Optional: tesseract (OCR), vosk or useful-moonshine-onnx (free offline transcripts for videos without captions).

Related MCP server: llm-video-mcp

Use

In your agent, just ask:

/glideit https://youtu.be/VIDEO_ID what happens at 12:30?

Or run the CLI directly:

# 1. MAP — whole-video transcript + storyboard grids
python scripts/glideit.py "https://youtu.be/VIDEO_ID"

# 2. ZOOM — dense high-res frames of one window (+ OCR of on-screen text)
python scripts/glideit.py "https://youtu.be/VIDEO_ID" --start 12:00 --end 13:30 --resolution 1024

The map prints paths to transcript.txt and storyboard_*.jpg; the zoom prints per-frame paths. The agent Reads those files and answers. Everything is cached under .glideit/<hash>/ — re-runs are instant.

Options

MCP server

mcp/server.py exposes map_video, zoom_video, and note_video to any MCP host (pip install mcp):

"glideit": { "command": "python", "args": ["mcp/server.py"] }

Recreate or remix a video

--cards turns a reference video into an editable template: a structured cards.json (per-scene text, narration, timing) plus a starter HyperFrames composition. Change the content, brand, or language and render a new MP4 — then run glideit on the render to review it. The demo video above was made this way.

How it compares

claude-video's /watch is great for short clips; glideit is built for the long ones — full lectures, tutorials, conference talks — plus OCR for on-screen code and the recreate/remix bridge.

License

MIT