YouTube MCP
that makes the video.

An MCP (Model Context Protocol) server gives an AI agent tools it can call. A YouTube MCP server lets Claude Code, Codex, Grok Build or any MCP client work with YouTube: search videos and Shorts, read a video's details and transcript and, in NotPeople's case, run the production work around it (cutting, captions, thumbnails, generated films) on your own machine. You add it with one line and ask in plain language.

Most YouTube MCP servers do one of two things: read data (search, stats, comments, transcripts through the YouTube Data API or yt-dlp) or publish through the API. NotPeople does the research part too, and then makes the video: it finds the moments in a long episode, cuts and reframes them into Shorts with burned captions, translates the captions, draws thumbnails, and renders animated, product and data films from code. Each job follows a written playbook, so the agent works in a fixed order and doesn't improvise.

No. The finished file stays on your disk and its title and description are saved as a draft for you to review. You upload. Publishing isn't part of the server today, which also means nothing reaches your channel without you seeing it first.

Shorts cut from your long videos or podcasts, 9:16 with captions in your style and in several languages; thumbnails; code-drawn animation from a 3-second loop to about a minute, at 1920×1080 60 fps plus native vertical, square and GIF versions; 30–50 second product films shot from real captures of a website; and short data-driven films built from real market or system data. Longer edits of your own footage go through a local timeline editor the agent drives over MCP. Generated films top out around a minute, and masters are 1080p.

Not built in. Films come with a synthesised soundtrack mastered to about −14 LUFS, and your own voice track drops onto the timeline like any other audio. If you use a text-to-speech service, the agent can place its output; the voice itself comes from you or that service.

ffmpeg and yt-dlp, which the setup playbook installs. Transcription runs on a local speech model. The timeline editor needs Node 24 and is comfortable from 16 GB of RAM; the agent asks before installing it. Animation and product films render on a CPU. 3D scenes and the fastest data-film path want an NVIDIA GPU. Without one, the agent can rent a GPU box on Vast.ai with your key, run the render and shut the box down; a ten-minute test render cost us $0.06.

No, though the idea is close: video as code, rendered frame by frame through a headless browser into ffmpeg, with a live preview you can scrub before the render. NotPeople's recipes use their own React/canvas scenes, Playwright captures and Blender for 3D.

Claude Code, Codex and Grok Build have one-line install commands on this page. Claude Desktop, Cursor and any MCP client that speaks HTTP transport work the same way.

The free tier covers the research tools and drafts. Creator ($49/mo) unlocks every playbook, including clips. Pro ($99/mo) adds the full video pipeline and 300 image credits. AI stills, AI video clips and rented GPUs run on your own keys (Gemini, Picsart, Vast.ai), so you pay those providers directly. No card for the free tier, cancel anytime.