# Chromuse

- **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:** [An Sung](https://cerebralvalley.ai/u/devansung)
- **GitHub:** https://github.com/dev-ansung/chromuse
- **Demo video:** https://www.youtube.com/watch?v=fgoRNn6s9ic
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5159

Chromuse turns the Chrome you already use into an assistant that remembers you and learns how to get things done. Your tabs. Your logins. Your browser. Give it a task by typing, speaking, or attaching an image, and watch it navigate, click, and work through the steps live.

Powered by Muse Spark Open, Chromuse is built around four strengths:

- **Fast:** No separate browser to launch or accounts to sign into again. It reads your open tabs before the first model call and trims stale page snapshots to keep the loop lean. Local telemetry shows a median first model response of 2.4 seconds across 74 tasks; whole browsing tasks took a median of 19.1 seconds across 54 tasks, on one machine.
- **Remembers:** Your preferences carry into future conversations through MEMORY.md. You can read, edit, and undo what it remembers.
- **Learns from experience:** A difficult task becomes a reusable playbook: the working steps, plus the pitfalls to avoid. A local embedding model finds relevant workflows, and the agent reads their steps only when needed.
- **Understands you:** SOUL.md lets you shape how it behaves. You and the agent can both edit its soul and memory, with version history and your changes taking precedence.

Muse Voice Transcribe makes speaking a task as natural as typing one. A live browser view, token usage, and local traces let you see what happened and debug what went wrong.

Built solo with LangChain, LangGraph, chrome-devtools-mcp, React, and assistant-ui. Today, Chromuse runs locally for one person, with prompt-based safety rules. The goal is a browser assistant that carries useful experience into the next task.

## More from Meta Global AI Developer Hackathon

- [FleetLLM](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5156)
- [AtMoS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5157)
- [Stevens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5158)
- [NEUROLEARN AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5160)
- [I surpass my yesterday self](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5161)
- [Access Denied](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5162)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5159. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt