macOS app that records meetings (microphone + system audio via ScreenCaptureKit), transcribes them with WhisperKit, and summarizes with a configurable LLM provider (Claude / Gemini / OpenAI / local Ollama).
Author: Alex Pavlyuk (@Alex-Pavlyuk)
cd MeetingRecorder
./build_app.sh
open dist/MacMeetSubscriber.appThe first build signs the app to establish TCC identity. Subsequent builds only replace the binary (preserving mic / screen capture permissions).
Open the app → Settings.
- LLM Provider — pick Claude / Gemini / OpenAI / Ollama
- API keys — per-provider, stored in macOS UserDefaults
- Ollama URL — default
http://localhost:11434 - Ollama model — auto-discovered from the running Ollama instance
- Context window (num_ctx) — see the memory section below
- Whisper model —
small(multilingual, recommended) ortiny(fast) - Use multi-stage pipeline — see pipeline section
- Auto-summarize after recording — run transcription + summarization automatically after Stop
- Start Recording → mic and system audio are captured to
mic.wav/system.wav - Stop Recording → app transcribes with WhisperKit (multi-pass language detection) and produces
transcript.md - If auto-summarize is on, the summarizer runs and writes
summary.md - Each recording lives under
~/Documents/MeetingRecorder/YYYY-MM-DD_HH-MM-SS/
You can also re-summarize existing recordings: right-click the row or use Summarize Last / Summarize All Pending.
The KV cache grows linearly with num_ctx. Pick the smallest value that still fits your meetings — larger contexts need more RAM and slow down prefill.
40-minute meeting ≈ 18-22K tokens of transcript. It will only fit completely when num_ctx ≥ 32768.
| Meeting length | Recommended num_ctx |
|---|---|
| ≤ 15 min | 8192 |
| 15 – 25 min | 16384 |
| 25 – 55 min | 32768 ← most common |
| 55 min – 1.5 h | 65536 |
num_ctx |
KV cache | Model | Total in RAM |
|---|---|---|---|
| 4096 | ~0.5 GB | ~5 GB | ~5.5 GB |
| 8192 | ~1 GB | ~5 GB | ~6 GB |
| 16384 | ~2 GB | ~5 GB | ~7-8 GB |
| 32768 | ~4 GB | ~5 GB | ~9-10 GB |
| 65536 | ~8 GB | ~5 GB | ~13-14 GB |
For 12B models (gemma3) add ~2-3 GB.
Symptom: tok/s is 10× lower than expected (e.g. 2-3 tok/s on an 8B model that should run at 20-30 tok/s on Apple Silicon).
Almost always the cause is memory pressure: several Ollama models loaded at once, causing swap. Check:
# Show currently loaded models and their size:
curl -s http://localhost:11434/api/ps | python3 -m json.tool
# Show swap activity; if Swapouts is growing, you're thrashing:
vm_stat | grep -E "free|Swapouts"Fix:
# Unload every model except the one you want:
for m in qwen3:8b qwen2.5:7b llama3.1:8b gemma3:12b deepseek-r1:8b mistral-nemo; do
curl -s http://localhost:11434/api/generate -d "{\"model\":\"$m\",\"keep_alive\":0}" > /dev/null
done
# Prevent this from happening again — put this in ~/.zshrc or your LaunchAgent:
export OLLAMA_MAX_LOADED_MODELS=1
# If Ollama has been running for days, restart it:
brew services restart ollama
# Keep Ollama up to date — older 0.20.x releases have Metal regressions:
brew upgrade ollamaOther causes of slow generation:
- Thermal throttling (common on MacBook Air / base M2 under sustained load — let it cool down)
- Other apps holding wired memory — close Chrome / Docker / Xcode if tight on RAM
- Model running on CPU (unusual). Check
size_vramin/api/ps— should equal model size
Settings → Summarization → Use multi-stage pipeline switches from a single LLM call to a 6-stage process:
- Chunk the transcript into 5-minute windows (no LLM)
- Extract structured JSON (topics, points, decisions, action items) per chunk — N LLM calls with JSON format
- Cluster chunks by normalized topic name (no LLM)
- Write a themed section per cluster — K LLM calls
- Consolidate action items across chunks — 1 LLM call
- Overview — 1 LLM call
Good fit:
- Cloud providers (Claude / OpenAI / Gemini) — extra calls are cheap, quality jumps noticeably
- Local Ollama only if generation speed ≥ 15 tok/s (see troubleshooting above)
- Long meetings (1+ hour) where single-shot can't hold the full transcript in context
Bad fit:
- Slow local inference (<10 tok/s) — each chunk takes many minutes, overall time balloons to 40+ min
- Short meetings (<15 min) — single-shot already fits everything
| Setup | Single-shot | Pipeline |
|---|---|---|
| Claude Sonnet | ~20 s | ~40-60 s |
| Ollama @ 25 tok/s | ~8 min | ~10-12 min |
| Ollama @ 15 tok/s | ~15 min | ~18-22 min |
| Ollama @ 3 tok/s (thrashing) | ~40 min | don't |
Pipeline eliminates most date/fact hallucinations because each chunk sees only a small slice of transcript, so the model can't invent things that weren't said.
~/Documents/MeetingRecorder/2026-04-16_15-33-19/
mic.wav # microphone capture
system.wav # system audio capture
transcript.md # speaker-labeled, timestamped transcript
summary.md # LLM summary
metadata.json # duration, segment count, languages, provider, status
metadata.json.status is one of pending_summary (transcribed but not summarized) or complete.
Copyright © 2026 Alex Pavlyuk.
Licensed under the Apache License, Version 2.0 — see LICENSE.