# AUTOKEREN

- **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:** [ajat sudrajat](https://cerebralvalley.ai/u/autokeren)
- **GitHub:** https://github.com/autokeren/kancil
- **Demo video:** https://www.youtube.com/watch?v=WfeQW7qQk6Q
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/110

Kancil is the first agent harness that lives natively in Facebook Messenger — a billion-user channel with zero existing agent frameworks. The agentic AI wave (autonomous coding agents, agent harnesses) lives entirely on Telegram and WhatsApp. Messenger has none. Not because it's hard — Telegram bots deploy in 5 minutes while Messenger requires Meta app review. Everyone follows the easy path; the billion-user channel stayed empty. Kancil fills it. Built for Indonesia — one of the world's largest Facebook markets (140M+ users, 64M+ small businesses) — where UMKM owners, students, and warung operators will never install a CLI or open a terminal. They already voice-note their friends all day. Kancil meets them exactly where they are. Send a text, a voice note, a photo, or a video — Kancil builds it. All verified live on a real Messenger page: • Voice note → Website: "bikin website kedai kopi Kopi Sore" → agent writes code, publishes live URL with menu, hours, WhatsApp checkout — in seconds • Voice note → Web browsing: "buka detik.com, apa berita utamanya?" → agent opens its own stealth browser, reads the page, summarizes headlines • Photo → Website: send a photo of a menu or screenshot → agent reads it and builds a site from the content • Video → Description: agent watches the visual sequence and embedded audio, describes what happens • Self-verification: after publishing a site, agent opens it in its own browser, takes a screenshot, and sends it to the user as proof • Persistent memory: per-user JSON storage, 85% context compaction (Muse summarizes old turns) • Contextual understanding: user says "ya coba buka" → agent remembers previous context and opens Google News instead of Google homepage • Zero hardcoded replies: even the quick acknowledgment is written by Muse in the user's language via a parallel intent-router request • Real-time tool progress: "Writing code… → Publishing site… → Opening website… → Taking screenshot…" — the user watches the agent work

## More from Meta Global AI Developer Hackathon

- [CO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/107)
- [NeuroLab AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/108)
- [Old Forge Technologies](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/109)
- [Proofline](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/111)
- [Cengkudu 800 MDPL](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/112)
- [Medina](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/113)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/110. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt