MeetMouse

A local-first, zero-telemetry real-time leadership coach for macOS. It listens to a live meeting, applies a facilitation rubric, and surfaces short coaching calls while the meeting is happening — shown in a small, screen-share-safe overlay.

Forked from anarlog (MIT), which already handles audio capture, on-device transcription, and bring-your-own-LLM including local models. The coaching layer is the new work.

Privacy: transcription is fully on-device and telemetry is zero. AI runs locally via Ollama (127.0.0.1) by default and local mode works with WiFi off. Users can explicitly enable Claude or OpenAI in Settings → AI using their own API key. Cloud mode sends the text and context needed by AI features directly to the selected provider; audio stays on-device. API usage is billed by that provider. Keys are stored in device-only macOS Keychain.

To enable cloud AI, choose a provider and model, paste an API key, optionally test the connection, accept text sharing, then click Save and enable. Use local AI stops future cloud requests. Provider changes apply to the next meeting and the next on-demand AI request. A connection failure leaves transcription and built-in coaching working.

See PLAN.md for the full build plan and findings.md for the Phase 0 recon.

Install (macOS)

Download the app: grab the latest MeetMouse-*.dmg from Releases, open it, and drag MeetMouse to Applications. It's signed and notarized, so it opens with a normal double-click.

First launch: the app bundles the Ollama runtime, so you don't install anything else — just click Download model once to fetch a local model (needs WiFi for that one download; everything after runs offline).

Build from source

Requires Xcode 16+ and XcodeGen (brew install xcodegen).

git clone https://github.com/noahdevkagan/meetmouse.git
cd meetmouse/MeetingCoach
xcodegen
xcodebuild -project MeetingCoach.xcodeproj -scheme MeetingCoach \
  -configuration Debug -derivedDataPath build build
open -n build/Build/Products/Debug/MeetMouse.app

For local LLM features when running a source build, either have Ollama running, or vendor the runtime into the bundle with ./scripts/vendor-ollama.sh.

Working with an AI coding agent? Point it at AGENTS.md — setup, build, test, and repo-specific gotchas in one file (Codex reads it automatically; Claude Code loads it via CLAUDE.md; an optional .mcp.json adds Xcode build tooling over MCP).

Maintainers: see DISTRIBUTION.md for cutting signed releases.

Use your transcripts with AI

Every meeting is saved as plain Markdown in ~/Documents/MeetingCoach/transcripts/ (the path is shown — and changeable — in Settings → General). No integration needed: point Claude, ChatGPT, Cursor, or any other AI tool at that folder and it can read everything.

• Filenames sort and glob: 2026-08-31T14-30_partner-sync_chad-anna.md — ISO date-time, then the meeting title and participants, slugified.
• Machine-readable metadata: each transcript has a .json sidecar (title, started_at, duration_min, participants, source, filename), and index.jsonl in the folder lists every meeting, one JSON object per line, append-only.
• The folder deliberately lives in ~/Documents, not ~/Library — macOS lets you grant AI tools access there. Transcripts recorded before this layout were copied in automatically; the originals stay where they were.

For deeper access (search across all sessions from an agent), the app also bundles an MCP server — see Settings → General → Agent access.

Status

• Phase 0 (Recon): DONE — see findings.md. Recommendation: build the coach in-process (plugin / Tauri-event consumer), not a sidecar, because the live transcript exists only as an in-process Tauri event stream.
• Phase 1 (Offline simulator): next — the de-risk step. Build the offline loop + backtest harness in simulator/ before touching audio.

Layout

PLAN.md         build plan
findings.md     Phase 0 recon output
rubrics/        swappable rubric configs (personal.yaml = example, default.yaml = generic)
simulator/      Phase 1: offline loop + backtest harness
coach/          the live coaching loop (in-process)
overlay/        Phase 3 overlay UI
anarlog/        fork (added as submodule/fork; not yet vendored)

Anarlog

The anarlog fork is not vendored into this repo yet. Add it as a submodule when starting Phase 2:

git submodule add https://github.com/fastrepl/anarlog.git anarlog