Launch videos built by an AI agent that writes code, not a video model

A promo demo (launchvideo.io, open-sourced as diggerhq/shipvideo) shows a serverless ‘director’ agent that turns a URL or product description into a 1080p MP4 launch video in roughly four minutes and ~100k tokens. The notable trick is that there is no video-generation model in the loop: the LLM writes the film as HTML/CSS/JS, then a headless Chromium page has its clocks (requestAnimationFrame, timers, Date, CSS/Web Animations) swapped for a virtual clock so every frame is a deterministic seek. Frames are piped as JPEGs into libx264 at 30fps. Because the film is code, re-running the same input renders the same output every time.

The architecture is lean and worth noting for the security-minded. It runs on a platform called OpenComputer as a single TypeScript agent with three tools (web_fetch, check_scene, render_video). Each job executes in a fresh, throwaway microVM (Amazon Linux 2023, arm64, 4 vCPU/8GB, Node 22) that installs Playwright’s Chromium and a static ffmpeg on first call, then is discarded. The agent holds no secrets: the web form mints a path-scoped Vercel Blob upload token valid for three hours per job, and the finished MP4 lands at a public Blob URL. The same session/event-stream API the CLI uses drives the progress display.

One accuracy flag for readers: the page credits ‘anthropic/claude-opus-5.5,’ but Anthropic ships no Opus 5.5 — its current flagship is Opus 4.8 — so the model name should be treated as unverified. This is a vendor showcase from Digger/OpenComputer rather than independent reporting, but the underlying idea (an LLM authoring deterministic, code-rendered video instead of calling a diffusion-style video model) is the genuinely interesting takeaway.