OffReco Setup & how it works
Just follow the first-run wizard. No fiddly configuration.
Download & open
Get the .dmg and move it to Applications. The wizard opens on first launch.
Grant permissions
Allow microphone and calendar. No screen-recording permission needed.
Prepare transcription (auto)
Python and ffmpeg are bundled. The first run prepares the transcription libraries automatically (internet required).
Recommended model auto-downloads
It downloads the model for your language with a progress bar. Ready right after.
Speaker labels (optional)
Want “who said what”? Add a free HuggingFace token (optional).
Create meeting minutes
OffReco never calls an AI itself. You generate minutes with your own ChatGPT or Claude, saved as minutes.md (and synced with your project). Pick the shortest path for the AI you use. Only the transcript is shared with the LLM — never the audio, and only when you press the button.
Claude Desktop (MCP)
- 1 In Settings → “Integrations,” turn on Claude Desktop and click “Copy JSON config.”
- 2 Paste it into claude_desktop_config.json and restart Claude Desktop.
- 3 Open a meeting and press “Create with Claude Desktop” in the Minutes tab — summarizing and saving happen automatically.
Claude (web) / ChatGPT
- 1 Press “Create with Claude” or “Create with ChatGPT” in the Minutes tab (the prompt + transcript are copied and your browser opens).
- 2 Paste with ⌘V → send → copy the minutes it returns.
- 3 Back in OffReco, a detection HUD appears — press “Import” to save it as minutes.md.
Three templates — Standard, Formal (internal), and Technical — are fully editable under Settings → Integrations, with {{transcript}}, {{date}}, {{duration}}, and {{attendees}} placeholders.
Choosing a model
For meetings, the model matters. On read-speech benchmarks the Japanese-tuned Kotoba is very accurate, but for natural meetings (overlap, fillers, noise) the general models (medium / large-v3) can be steadier. Apple Silicon runs them fast on the GPU (mlx); Intel runs on CPU. Pick what fits your Mac below (change it anytime under Settings → Transcription).
Short answer
Balanced for natural conversation, and fast on the GPU.
A larger, more robust model. 16GB+ recommended.
Japanese-specialized, light, runs on CPU. Strong on read-style speech.
Recommended for meetings, by Mac
Apple Silicon is fast on the GPU (mlx). Intel has no mlx GPU and runs on CPU, so lighter models are the safer bet. The Japanese-tuned Kotoba is light and runs on any Mac.
| Your Mac | For meetings | Lightweight / Japanese-tuned |
|---|---|---|
| Apple Silicon · 16GB+ (M1 Pro/Max, M2/M3/M4, etc.) | medium or large-v3 (mlx / GPU) | Kotoba Whisper v2.0 |
| Apple Silicon · 8GB (base M1/M2/M3) | medium (mlx / GPU) | Kotoba Whisper v2.0 |
| Intel · 16GB+ | Kotoba Whisper v2.0 / medium (faster · CPU) | small |
| Intel · 8GB | Kotoba Whisper v2.0 / medium (faster · CPU) | small |
| Lightweight / draft use | small | small |
When a meeting ends, transcription starts automatically. It can take a little while — that’s your Mac doing the work. Processing time depends on your Mac and the model you picked above (lighter models are faster; higher-accuracy ones take longer). You’ll get a notification when it’s ready.
All models
| Model | Languages | Accuracy | Speed | Size | Min RAM | Notes |
|---|---|---|---|---|---|---|
| Whisper large-v3 | Multilingual | ◎ | △ | ~3GB | 8GB+ (16GB+ recommended) | Top accuracy; robust on noise and natural conversation. mlx (GPU) on Apple Silicon, faster (CPU, slow) on Intel. The main pick for meetings. |
| Whisper medium | Multilingual | ○ | ○ | ~1.5GB | 4GB+ | Balanced. Fast via mlx on Apple Silicon — a handy default for meetings. |
| Kotoba Whisper v2.0 | Japanese | ○ | ◎ | ~1.5GB | 4GB+ | Japanese-tuned, light and fast (faster-whisper / CPU; no mlx build). Especially strong on Japanese TV-style audio and faster than large-v3. On Apple Silicon, mlx large-v3 / medium are faster and often steadier for meetings, so Kotoba is mainly for Intel or a light Japanese-tuned option. |
| Kotoba Whisper Bilingual v1.0 | Japanese · English | ○ | ◎ | ~1.5GB | 4GB+ | Handles both Japanese and English (faster-whisper / CPU). |
| Distil-Whisper large-v3 | English | ◎ | ◎ | ~1.5GB | 4GB+ | English-only. Within ~1% WER of large-v3, lighter and faster, with fewer hallucinations. |
| Whisper small | Multilingual | △ | ◎ | ~0.5GB | 2GB+ | Lightest & fastest. For low-spec Macs or drafts. |
Accuracy and speed are relative guides (◎ > ○ > △). Accuracy is a guide for read-style speech; for natural meetings (overlap, noise) results differ — the general medium/large-v3 can be steadier. Speed depends on your Mac, model size and engine (mlx = Apple Silicon GPU / faster = CPU).
When in doubt, start with medium (mlx) on Apple Silicon, or Kotoba Whisper v2.0 on Intel. The app automatically warns you if a model won’t run on your Mac (not enough RAM, or Apple Silicon-only).
Try it free
First month free. Install and follow the first-run wizard — that’s it.