Tractus
Free and open source · macOS · Apple Silicon
Tractus records your meetings, transcribes them on-device in Korean and English, tells speakers apart, and turns them into summaries you can search, cite — and ask Claude Code about.
Rough memo → enhanced note with details from the transcript → ask Claude Code about the meeting (sample data).
Why Tractus
Recording, speech recognition, speaker detection and smart search run on your Mac with Apple MLX. No audio is uploaded, and none of it needs an account or API key.
Each summary item, to-do and AI answer links to the moment in the recording it came from. Click to listen.
Notes live in a folder as audio, JSON and Markdown. Export to TXT, Markdown, SRT, VTT, CSV, DOCX or PDF, or send a note to Obsidian.
A built-in MCP server and a one-line Claude Code plugin let your agent search your meetings and answer with citations like [Weekly dev sync 00:02:11].
Features
Record in person or on video calls, then read, fix and reuse what was said.
Qwen3-ASR transcribes Korean and English on your Mac with automatic language detection. Speakers are detected and can be named once — Tractus suggests the name in later recordings.
Meeting notes, brief, to-dos, lecture, interview or your own instructions. Ask one note or all of them — every answer cites the lines it's based on.
Memos with checklists and formatting, favorites, tags and folders, and smart search that matches words and similar meanings across every note (⌘K).
Measured, not guessed
We benchmarked on-device models on real Korean meetings (5.2 hours), code-switched speech and English meetings on an M3 Mac. Lower is better.
| Model | Korean meetings (CER) | Korean + English (MER) | English meetings (WER) | 1 hour of audio |
|---|---|---|---|---|
| Qwen3-ASR-1.7B (Tractus default) | 9.49 | 24.1 | 9.50 | 8.2 min |
| Qwen3-ASR-0.6B (lite) | 11.47 | 26.1 | 10.14 | 3.3 min |
| Whisper-large-v3-turbo | 14.92 | 26.6 | 14.28 | 6.0 min |
| Nemotron 3.5 ASR 0.6B | 10.93 | 45.1 | 14.13 | 7.6 min |
AI Hub meeting speech, HiKE and AMI, MLX 8-bit, Apple M3 16 GB. Read the full benchmark →
For Claude Code
Install the plugin and Claude Code can search your meetings, read transcripts and to-dos, and cite the exact moment. Your notes stay on your Mac; only what Claude reads is sent to Claude.
# in Claude Code /plugin install tractus --marketplace sjoon21/Tractus > When did we decide to ship the release? The release moved to October 22 — a week later than planned, because of the payment API issue and too little QA time. [Weekly dev sync 00:02:11]
FAQ
Yes. Tractus is free and open source under the Apache 2.0 license. Recording, transcription, speaker detection, search and export work without any account or API key. Summaries and Ask AI use an AI service you choose (Upstage, OpenAI, Claude, Gemini or any OpenAI-compatible API, including one on your Mac).
No. Speech recognition, speaker detection and smart search run on your Mac with Apple MLX. Only the text needed for a summary or an AI answer is sent, and only to the AI service you pick.
Apple Silicon Macs (M1 or later) on macOS 14.4 or later. 16 GB of memory is recommended; 8 GB works with the lite models. You also need ffmpeg (brew install ffmpeg).
Tractus detects the language automatically and is tuned for Korean and English meetings, including Korean speech mixed with English technical terms. The app itself is available in English and Korean.
Tractus covers the same core workflow — record, transcribe, summarize, search — but runs the transcription on your Mac instead of a cloud service, and keeps every note as plain files you own.
Not yet. On first launch, if macOS says the developer can't be verified, open System Settings › Privacy & Security and click Open Anyway. See the setup guide.