Built ambient voice intelligence for AI agents using Omi 🦞

[davidemanuelDEV]

Hey! Wanted to share a project we built on top of Omi.

Percept is an open source pipeline that turns your Omi pendant into an ambient intelligence layer for AI agents. Built for OpenClaw 🦞 but designed to work with any agent framework.

It goes beyond simple transcription — it understands context, identifies speakers, extracts entities, and routes voice commands to your agent.

Architecture:
Omi pendant → phone (STT) → transcript webhook → Percept receiver → wake word detection → intent parsing → OpenClaw 🦞 agent execution → conversation summaries → entity extraction → knowledge graph

What makes it different from basic webhook integrations:
• Wake word system — only triggers agent actions on specific phrases
• Speaker authorization — only approved speakers can trigger commands
• Context Intelligence Layer — knowledge graph across conversations
• Continuation windows — handles multi-sentence commands naturally
• Auto-summaries — meeting notes when conversations end

We also published 5 skills on ClawHub (OpenClaw's skill marketplace):
🎧 percept-listen — stream live transcripts
📝 percept-summarize — auto meeting summaries
🎤 percept-voice-cmd — wake word → action
🗣️ percept-speaker-id — know who's talking
🌍 percept-ambient — always-on intelligence

What we learned building on Omi's webhook API:

1. Transcript webhooks > audio byte webhooks (Omi's STT is solid)
2. The is_user flag is invaluable for speaker identification
3. Segments arrive in bursts — need buffering logic with silence detection
4. GET validation endpoint needed for webhook registration

GitHub: github.com/GetPercept/percept
License: MIT
Works with: OpenClaw 🦞 (and extensible to any agent framework via CLI)

Would love feedback from anyone building agent integrations on Omi. PRs welcome — first 50 contributors get lifetime Pro access when our hosted API launches 🙏