XooMe is the open-source Mac app that records your meetings, listens to your day via the pendant you already own, and turns the whole mess into searchable transcripts, summaries, and to-dos. Bring your own keys. Run it 100% local if you're paranoid. We respect that.
Thousands of us bought a wearable BLE pendant for always-on capture. Then the company was acquired and the service went dark. The hardware kept working. The protocol just needed someone to read it. So we did.
XooMe is the result — a reverse-engineered, open-source macOS companion app that talks to your existing pendant over Bluetooth, syncs the audio off the onboard NAND, and turns it into something you can actually use. Works with the original hardware. Designed to support more.
XooMe runs the same transcription, diarization, and AI pipeline behind both modes — but the products are different on purpose. Use what you need. Ignore the other.
Hit record. XooMe captures system audio + your mic, transcribes live with speaker diarization, and produces a clean summary plus action items when you stop.
Wear the pendant. Forget about it. XooMe detects encounters from the day's audio, clusters speakers across days, and gives you a queryable timeline you actually want to revisit.
SwiftUI, NavigationSplitView, real keyboard shortcuts. The pendant lives in your menu bar. Both feel like Mac apps because they are.
XooMe ships with no backend, no telemetry, no account. Point it at a frontier API or your own self-hosted endpoint — the app doesn't care, and neither does your wallet.
Audio, transcripts, notes, and summaries are JSON-per-record under ~/Library/Application Support/XooMe/. Backup is a folder copy.
Deepgram for transcription. Anthropic, OpenAI, or any compatible endpoint for AI. No middleman markup. No "credits."
Point the AI dispatcher at a local model — Ollama, llama.cpp, your DGX cluster, whatever. Same UI, zero packets leaving your network.
Drop your keys into Settings. Switch providers any time. Or skip the cloud entirely and route everything to a local endpoint. The escape hatch is part of the architecture, not an afterthought.
Real-time WebSocket streaming during a meeting. REST batch processing for pendant audio after the fact.
Server-side speaker turns from Deepgram. In Lifelog, voiceprint clustering keeps speakers consistent across days.
Claude (or your model) generates summaries and structured action items grouped by person and by team.
VAD-bounded sessions get clustered into actual conversations — no more 30 fragments of one car ride.
Coming next: a chat panel that RAGs over your full transcript corpus. "What did Richard say about pricing in March?"
Calendar auto-record, Obsidian sync, push transcripts as context to ChatGPT, Claude, or Cursor.
This isn't a roadmap PDF. It's the actual project board, lightly edited so it reads like English.
The whole stack — Swift app, BLE pendant client, build & sign scripts, the protocol notes — is on GitHub. Issues open. PRs welcome. Pendant hardware support is built to extend.
