LaunchVideo (launchvideo.io)

Live: https://launchvideo.io

Paste a URL or a prompt, get a 20 to 40 second launch video. No video model: Opus 5.5 writes a single HTML film and a serverless agent on OpenComputer renders it frame by frame in headless Chromium, encodes it with ffmpeg, and uploads the MP4 to Vercel Blob. Same mechanism as the "Opus 5.5 is incredible at instructional video generation" demo (Opus wrote JS, rendered via a headless browser + ffmpeg), packaged behind a one-field form.

opencomputer/   the OpenComputer project (one agent, `director`, model anthropic/claude-opus-5.5)
oc-template.toml  one-click deploy manifest
web/            Next.js app: the form, two API routes, Vercel Blob for storage

One click: https://app.opencomputer.dev/new?repository-url=https%3A%2F%2Fgithub.com%2Fdiggerhq%2Fshipvideo

Or from a terminal:

npx opencomputer template deploy https://github.com/diggerhq/shipvideo

The agent works on its own in the playground or the CLI (it keeps the MP4 in the runtime and tells you the path). The web/ frontend adds the form and the upload to Vercel Blob.

1. POST /api/jobs validates the input, mints a job id, generates a Vercel Blob client token scoped to videos/<jobId>.mp4 (3 h, video/mp4 only), creates an OpenComputer session, waits for runtime.connected, and sends one turn whose text is a JOB block (mode, input, upload pathname, upload token).
2. The agent (url mode) fetches the page with web_fetch, which also extracts title, description, headings, the most used hex colors, and Google Fonts.
3. It writes the film as one HTML document (1920x1080, CSS keyframes or a rAF loop), runs check_scene (loads it under the virtual clock, reports JS errors and the visible text at several timestamps), fixes, then calls render_video.
4. render_video renders every frame with Playwright's arm64 headless shell under an injected virtual clock (rAF, timers, Date, and CSS/WAAPI animations are all driven by __seek(t)), pipes JPEG frames into a static ffmpeg (libx264, crf 18, yuv420p, faststart), and uploads with @vercel/blob/client.
5. GET /api/jobs/<id>?session=<sid> reads the session's events to derive a phase, and reports done as soon as head(videos/<id>.mp4) finds the blob.

The runtime is a fresh Amazon Linux 2023 arm64 microVM per session (node 22, dnf, no browser), so the first tool call installs playwright-core, ffmpeg-static, @vercel/blob, the Chromium shared libs, and Noto fonts into ~/workspace/renderer (about a minute). Rendering runs at roughly real time: a 30 s film takes 30-40 s.

npm install && npx opencomputer link --create-project shipvideo && npx opencomputer deploy --watch
cd web && npm install && vercel link && vercel env pull .env.local
# add to web/.env.local: OPENCOMPUTER_API_KEY=... and OC_AGENT_ID=shipvideo@development
npx opencomputer env set BLOB_PUBLIC_BASE --value-stdin   # https://<store>.public.blob.vercel-storage.com
cd web && npm run dev

The web app needs a Vercel Blob store connected to the project (public access) so BLOB_READ_WRITE_TOKEN lands in the env. Nothing else is required: the agent has no secrets, the only credential it ever sees is the per-job upload token.

• Turns carry a plain string (POST /sessions/:id/turns {input}), so the job fields ride in the prompt text and the model copies them into render_video.
• No <video>, <audio>, <iframe>, CSS transitions, Math.random, or external images in a scene; the tools reject the first three and the prompt forbids the rest so renders stay deterministic.
• @sparticuz/chromium does not ship arm64 builds; Playwright's chromium-headless-shell does, and needs the dnf libs listed in opencomputer/agents/director/tools/renderer.ts.