# Excalibur agent

- **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:** [Cristhian Rudolf Egoavil Huaman](https://cerebralvalley.ai/u/cristhianr7)
- **GitHub:** https://github.com/CristhianRudolf7/excalibur-agent.git
- **Demo video:** https://youtu.be/7_hBhqf6aok
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4142

Excalibur is a personal AI agent that doesn't just talk. It acts: in your day, in your code and in the physical world. It is a single Go binary with a web interface, for Linux, Windows and macOS.

Your day: talk to it in a live voice call. It operates a real browser while you watch every step, and it fills passwords from a vault the model never sees. It remembers past conversations, cites the exact page of your documents, reads your Gmail and asks before sending. It searches the web with sources, runs automations and creates slides, Word and Excel files.

Your data: give it a messy spreadsheet and it proposes a clean database with linked tables, plus your own interface to view and edit it.

Your code: create as many agents as you need, each with its own chat. An agent reads a spec, splits the work among parallel subagents, tests its own code and sends you a public Cloudflare link to the result.

The physical world: the Robotics Lab is a harness extension that shows how Excalibur adapts to scientific needs. In simulation, a language model looks through the cameras and decides every joint of a robot arm and a robot dog. When the dog falls, the agent writes down why and tries again. In the run shown in the video, starting from a stair-climbing recipe, it reached the top on attempt 5.

Under the hood: a master agent coordinates subagents, tools and extensions, with a context compiler keeping each model call focused.

Measured, not promised: on Terminal-Bench 2.1, Excalibur scored 59.6% (53 of 89 tasks), ahead of OpenCode (55.1%), Terminus 2 (51.7%) and Codex (48.3%), using the same model for every harness (GPT-6 Luna, low reasoning effort), with one attempt per task, for about $1.55 for the whole benchmark.

## More from Meta Global AI Developer Hackathon

- [Provenance](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4139)
- [INVICTUS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4140)
- [Leonardo Rodriguez](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4141)
- [Araav Nayak](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4143)
- [Apni Gati](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4144)
- [Bypass Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4145)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4142. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt