# Ode: Give Muse a body

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Joe Hu](https://cerebralvalley.ai/u/Joe_Hu)
- **GitHub:** https://github.com/hubeiqiao/ode-muse-body
- **Demo video:** https://www.youtube.com/watch?v=0wcoqi5CVuQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2598

The idea.** Most AI agents live in a browser tab or behind an API. We gave Muse a body instead: a spare iPad mini that belongs to the AI full time. It has its own apps, its own memory and a name, Ode. You talk to it live, it can see you through the camera, and it operates the iPad the way a person does.

**How it works.**
- **Eyes:** the iPad mirrors its screen over AirPlay to a Mac at native resolution. Local OCR reads the text.
- **Brain:** Meta's Muse Spark (via the Meta Model API) looks at each screenshot, decides one action at a time and explains its reasoning. A judge model approves each step before anything happens, and a backup Muse judge takes over when the judge is unavailable.
- **Hands:** an ESP32 board pairs with the iPad as a Bluetooth keyboard and mouse. It types, presses shortcuts and clicks at calibrated screen coordinates. No jailbreak, no Xcode, no app APIs, so it works with apps that have no API at all.
- **Voice:** a live voice companion (Gemini Live) handles conversation and camera vision, and hands device actions to Muse Spark.

**What it can do today (verified on a real iPad in landscape).** Open and navigate Settings, YouTube, Instagram, Discord, X, LinkedIn, Grok, ChatGPT, Chrome and Weather from a spoken request; open the camera and describe what it sees; search Amazon; draft replies and read them back. Anything public waits for an explicit spoken "post it".

**Hard parts we solved.**
- **Landscape mirroring:** the AirPlay receiver advertised 16:9 by default, so iPadOS re-laid out some apps for the external screen and the AI saw one layout while clicking another. We advertise the iPad's native resolution and scale frames back for the vision pipeline.
- **Invisible pointers:** many apps hide the pointer in the mirror. When the pointer can't be seen, taps fall back to a freshly re-homed, calibrated path and verify that the screen changed.
- **Safety:** the device is the AI's own, but purchases, logins, passwords and settings changes are refused, and public posts need confirmation.

**What's next.** Turning any idle phone or tablet into a personal AI body: one ESP32 and a spare device, no special hardware.

## More from Meta Global AI Developer Hackathon

- [Jeffery Nketiah](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2595)
- [Fight Club](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2596)
- [SignBridge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2597)
- [The Gatekeepers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2599)
- [NOVA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2600)
- [Hull Forge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2601)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2598. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt