An agent skill + CLI + MCP server that turns one-sentence requests into finished videos rendered entirely on your machine. Motion graphics are HTML/CSS/canvas captured frame-exactly in headless Chrome; math scenes use Manim; footage is cut by transcript with ffmpeg. No API keys, no uploads — models (Kokoro/Piper voices, Parakeet transcription) live in ~/.showtime, about 1.4–1.8 GB unpacked, plus ~490 MB for the transcription model when first needed. No GPU required; WebGL looks render on the CPU. The engineering discipline is what stands out. Fixes re-render only the changed seconds (--from/--to). Images stay out of the agent's context via a disposable text reviewer with a 12-image-per-job budget. Every job ends with a receipt, a critic round by default, and pairwise blind review where a new render ships only if two critics both prefer it. The behaviour scoreboard (35/42 with the skill vs 25/42 without, rerunnable for $36) is the right kind of claim: narrow, reproducible, honest about what it doesn't cover. Tested end to end on macOS, Linux, and Windows runners; Windows 10/11 desktop x64 explicitly not yet run. Tested hosts: Claude Code, Codex, Cursor, Devin, OpenCode; everything else is "should work." Satya will put a Copilot key on the clapboard by Thursday, and Jensen will note that the more frames you render, the more you save.