Prompt → LLM-written Manim script → sandboxed render → video. Multi-scene, with per-scene refinement ("make the rectangle blue") that re-renders only that scene.

The Hono API validates a prompt, writes a Project row, enqueues a BullMQ job, and returns 202. A Bun worker plans scenes with Claude, generates Manim code per scene, renders each scene inside a locked-down Docker container (--network=none, CPU/memory caps, timeout), feeds render stderr back to the LLM for up to MAX_FIX_ATTEMPTS self-corrections, then concatenates clips with FFmpeg. The Next.js app polls status and shows per-scene previews with a refine box. LLM-generated code is untrusted — it only ever executes inside the sandbox container.

• Bun 1.2+
• Docker (for Postgres, Redis, and the Manim sandbox)
• FFmpeg + ffprobe on the host (used for concat + duration probing)

bun install
cp .env.example .env            # add your ANTHROPIC_API_KEY
docker compose up -d            # postgres + redis
docker pull manimcommunity/manim:v0.19.0   # pre-pull the sandbox image
bun run db:migrate              # answer the prisma prompts (name: init)

Milestone first — prove the pipeline with no servers:

bun run pipeline:test "a blue rectangle appears, then morphs into a circle"
# → prints the path to an MP4 when it works

Then run the whole app:

bun run dev                     # web :3000 · api :4000 · worker

• apps/web — Next.js UI: prompt in, scene timeline + previews, refine per scene
• apps/api — Hono: validate → persist → enqueue → 202; serves clips/videos
• apps/worker — BullMQ consumers: generate-video + scene-refine pipelines
• packages/pipeline — codegen (plan/generate/fix/refine prompts), Docker sandbox runner, FFmpeg concat
• packages/database — Prisma schema + client singleton
• packages/queue — queue names, connection, typed job payloads
• packages/shared — API DTO types shared by web + api

• Single-host v1: api + worker share storage/ on one machine. The worker shells out to docker, so in production either mount the Docker socket or swap sandbox.service.ts for Modal/E2B.
• No auth yet (deliberate, to reach the magic moment fast). Add Hono JWT middleware + a User model when you need it.
• Later: xfade transitions, voiceover via manim-voiceover (extend sandbox/Dockerfile), scene reordering, WebSocket progress instead of polling.