🎬 Bee One Animation Studio

Turn any story — a novel excerpt, a training dialog, a corporate policy, a bedtime story — into a playable 1980s Saturday-morning cartoon, rendered live in your browser with real character voices.

Bee One is a single local Python app that serves three things on one port:

Your LLM agent (Claude Desktop, Claude Code, Cursor, etc.) connects over MCP, reads the authoring guide, turns your story into a JSON action script, validates it, and pushes it straight onto the stage.

Install & run

pip install -e .
bee-one            # optionally: bee-one --port 9000

Open http://localhost:8888 in Chrome or Edge (they ship the natural "Online/Neural" voices) and click START STAGE — the click is what grants the browser permission for sound and speech.

Connect an LLM client

Streamable HTTP (recommended) — add an MCP server with URL http://localhost:8888/mcp. For example, in Google Antigravity (agy), edit ~/.gemini/config/mcp_config.json:

{
  "mcpServers": {
    "bee-one": {
      "url": "http://localhost:8888/mcp"
    }
  }
}

then restart the server list with /mcp inside agy. (The key — "bee-one" here — is the name your client shows for the tools.)

stdio (fallback) — for clients that can't do HTTP MCP: command bee-one, args ["--stdio"]. Your client launches this and talks JSON-RPC over its stdin — running it by hand in a terminal just waits forever. It relays scripts to the running bee-one web server, so start that first.

Then just ask your agent:

Here's a short story: … Turn it into a cartoon on my Bee One stage.

MCP tools

The schema is defined once in src/bee_one/schema.py (Pydantic) and the guide is generated from the same registry the engine implements — docs can't drift from reality.

What the engine can do

• Cel-look rendering: flat fills, chunky ink outlines, limited retro palette, characters animated on twos (12 fps poses at 60 fps playback), iris-wipe scene transitions, title cards, optional CRT scanline overlay.
• Three rig types: parameterized people (skin/build/hair/hat/glasses, 8 expressions, 10 gestures, held items, real walk/run/sneak cycles with bending knees), robots (treads, claw arms, LED equalizer mouth), and cars (5 styles, wheels that actually roll, exhaust, night headlights).
• Six layered backgrounds (diner, city_street, living_room, office, lab, forest) with day/sunset/night variants, weather (rain/storm/snow), ambient life (neon flicker, passing traffic, sleeping cat, bubbling beakers, fireflies…) and true depth: actors pass behind counters and trees and in front of walls.
• Voices & sound: Web Speech API TTS with per-character natural-voice assignment and live lip flap, plus a procedural WebAudio retro sound bank (boom, zap, boing, sad trombone…). No audio assets, no subscriptions.
• A cue-graph timeline: cues run in sequence by default; actions inside a cue run in parallel; after + delay create overlaps (talk while walking while the camera tracks you).

Try it without an LLM

Press ▶ PLAY DEMO on the stage — it plays static/demo.json, a three-scene episode generated from short_story.txt.

You can also drive it with curl:

curl -X POST http://localhost:8888/api/play -H "Content-Type: application/json" -d @src/bee_one/static/demo.json

Notes & limits

• Keep the stage tab focused/visible — browsers throttle timers and speech in background tabs. Best under ~10 minutes per script; the validator warns on longer estimates.
• Voice quality depends on the browser. Edge on Windows has the best free natural voices; everything still works (with plainer voices) elsewhere, and with no voices at all the show plays silently with speech bubbles.
• If your agent plays a script while no stage tab is open (or before you click START STAGE), the server remembers it: the next tab to open gets it queued, and START STAGE (or ⟳ replay) plays it. ▶ PLAY DEMO always plays the bundled demo — it does not play your agent's script.