What it does
Records the call, with a strip that stays out of the way
The app notices when a conversation starts and records it, or you press Record. The microphone and the system audio go to separate channels, so your voice is never confused with the people on the call. While it records, a strip under the menu bar asks you to tell the others, then shows the latest line it heard, with Pause and Stop at hand. The strip stays over full-screen apps and is left out of screen sharing.
Writes the transcript and learns the voices
Whisper turns the recording into a transcript; speaker separation splits the other side of the call into voices. Name a voice once and later meetings use the name when the voice matches. Click any line to hear it.
Keeps what was decided and who does what
With AI switched on, each meeting gets an overview, chapters, decisions, open questions and tasks with owners and deadlines. Today collects what happened recently and what is still open; Tasks lists every promise across meetings.
Follows a project across meetings
File meetings into a project and it keeps a living document: where things stand, the work list, decisions, questions and people, each linked to the meeting it came from. An item you edit is pinned: the model can mark it done but never rewrites it.
Answers questions from your archive
Search by words, or by meaning when AI search is on. Ask answers from your meetings, notes and projects and links every answer back to the passages it used.
Lets your AI tools read it
A read-only MCP server is built in. Claude Desktop, Codex, ChatGPT and VS Code with GitHub Copilot can search transcripts, open meetings, list tasks and read project state. See MCP access.
Everything else
Private by design
AI is optional. Without it, recording, transcripts, speaker names and keyword search all work. With the local option, summaries and answers come from a model that runs on the Mac, so nothing leaves it.
The MCP server only listens on 127.0.0.1 and never returns API keys or raw
voiceprints. The app asks GitHub once an hour whether a newer version exists
(Settings → Storage turns that off) and offers to install it; nothing else
leaves the Mac without your choice.
Install
- Download the
.dmgfrom the latest release. - Open it and drag Meeting Transcriber to Applications.
- Open the app. It is not notarized by Apple, so the first time macOS refuses to open it. Go to System Settings → Privacy & Security, find the message about Meeting Transcriber and choose Open Anyway. You do this once.
- Allow Microphone and System Audio Recording when macOS asks.
On the first launch the app downloads about 600 MiB into ~/MeetingTranscriber:
the transcription, speech-detection and speaker models, and FFmpeg.
The window opens straight away and shows the download; after that, recording
and transcription work offline.
Transcription uses Whisper large-v3-turbo, so it handles the languages Whisper does. Set the language in Settings or let it be detected for each meeting. English and Ukrainian are the most tested.
AI options
Choose in Settings → AI and archive. Recording and transcription do not depend on it.
- OpenAI — paste an API key.
- GitHub Copilot — the app fetches the Copilot CLI (about 90 MB), you sign in with GitHub in the browser and pick one of your plan's models. Requests use your plan's AI Credits; Business and Enterprise plans need the Copilot CLI policy enabled.
- Local — the app fetches Gemma 4 E2B (2.8 GB) and runs it with the bundled
llama-server. Paste a Hugging Face.gguflink to use a bigger model. Search by meaning can run locally too, with Qwen3 Embedding 0.6B (0.6 GB).
Where your data lives
~/MeetingTranscriber/
├── config.toml settings, the AI choice, and an optional OpenAI key
├── meetings.db transcripts, summaries, projects, notes, and search data
├── recordings/ captured and imported audio
├── models/ downloaded models, local AI models included
├── bin/ FFmpeg, and the Copilot CLI when chosen
├── copilot/ the Copilot CLI's own state
├── cache/ playback files
└── logs/mt.log the application log
MT_HOME moves the whole folder. Audio is kept for 30 days by default; deleting
audio never deletes the transcript or anything made from it.
MCP access
The MCP server is part of the app, not a second program. It opens as soon as the database is ready, without waiting for the speech models.
flowchart TB
DB[("The same SQLite archive<br/>the app uses")]
Tools["9 read-only MCP tools"]
HTTP["Streamable HTTP<br/>127.0.0.1:8765/mcp"]
Stdio["stdio mode<br/>--mcp-stdio"]
Codex["Codex / ChatGPT"]
VSCode["VS Code<br/>GitHub Copilot"]
Claude["Claude Desktop"]
DB --> Tools
Tools --> HTTP
Tools --> Stdio
HTTP --> Codex
HTTP --> VSCode
Stdio --> Claude
classDef data fill:#ECFDF5,stroke:#10B981,color:#064E3B
classDef transport fill:#EEF2FF,stroke:#6366F1,color:#1E1B4B
classDef client fill:#FFF7ED,stroke:#F97316,color:#7C2D12
class DB,Tools data
class HTTP,Stdio transport
class Codex,VSCode,Claude client
The tools list and open recordings, search transcripts and stored knowledge, read projects, list action items, build a recent briefing and list recognised people. Codex, ChatGPT and VS Code connect to the HTTP endpoint while the app is open; Claude Desktop starts the same executable in stdio mode.
Settings → MCP Server shows the live status and ready-to-copy configuration for each client. The same examples are in docs/mcp.md.
How it works
One process owns everything: the React interface is embedded in the Go executable by Wails, and the same process records, runs the queue, keeps the SQLite database and serves MCP.
flowchart LR
Person["Person at the Mac"]
Clients["Claude · Codex · GitHub Copilot"]
AI["OpenAI · GitHub Copilot<br/>optional, or a local model"]
subgraph App["Meeting Transcriber.app"]
direction TB
UI["React interface<br/>Wails webview"]
Core["Go application services"]
Capture["Microphone + system audio"]
Pipeline["Transcription + speakers"]
MCP["Read-only MCP server"]
UI <--> Core
Capture --> Core
Core --> Pipeline
Core <--> MCP
end
subgraph Local["~/MeetingTranscriber"]
DB[("SQLite archive")]
Audio["Recordings"]
Models["Model files"]
end
Person --> UI
Clients <--> MCP
Pipeline <--> Models
Core <--> DB
Capture --> Audio
Pipeline --> Audio
Core -. "summaries, embeddings, Ask" .-> AI
classDef app fill:#EEF2FF,stroke:#6366F1,color:#1E1B4B
classDef local fill:#ECFDF5,stroke:#10B981,color:#064E3B
classDef external fill:#FFF7ED,stroke:#F97316,color:#7C2D12
class UI,Core,Capture,Pipeline,MCP app
class DB,Audio,Models local
class Person,Clients,AI external
A live call and an imported file enter the same queue. A recording in progress always comes first, so old imports never make the current meeting lag.
flowchart LR
Start{"Live call or<br/>imported file?"}
Capture["Capture or copy audio"]
Queue["Add to SQLite queue"]
Decode["Decode and align channels"]
ASR["Transcribe locally"]
Speakers["Separate speakers"]
Names["Apply learned names"]
Save["Save transcript + voice evidence"]
Index["Build searchable passages"]
Summary{"AI chosen?"}
AI["Summarise and<br/>update the project"]
Ready["Ready in Library,<br/>Today, Search, and MCP"]
Start --> Capture --> Queue --> Decode --> ASR --> Speakers --> Names --> Save --> Index --> Summary
Summary -->|"yes"| AI --> Ready
Summary -->|"no"| Ready
classDef local fill:#ECFDF5,stroke:#10B981,color:#064E3B
classDef decision fill:#FFF7ED,stroke:#F97316,color:#7C2D12
classDef result fill:#EEF2FF,stroke:#6366F1,color:#1E1B4B
class Capture,Queue,Decode,ASR,Speakers,Names,Save,Index local
class Start,Summary decision
class AI,Ready result
The microphone channel is the person at the Mac; the system channel is everyone
else. Speaker separation runs on the system channel, so your own voice echoing
in the room does not become a second speaker. The .app carries the speaker
libraries, an audiotee helper for system audio and llama-server for local
models; the models themselves are downloaded per user.
The reasoning behind these choices, with the measurements, is in engineering decisions and architecture.
Build from source
You need an Apple Silicon Mac with macOS 14.2 or newer, Go 1.26.2, Node.js with npm, CMake and the Xcode Command Line Tools. The first build fetches the pinned native sources and frontend dependencies.
make # build/mt, the development binary
make bundle # build/Meeting Transcriber.app
make run # open the bundle, so macOS gives the permissions to the app
make install # replace /Applications/Meeting Transcriber.app with this build
make dmg # build/MeetingTranscriber.dmg
make test # go vet and go test
Always run the .app. Started as a bare build/mt from a terminal, the
microphone permission belongs to the terminal and the app records silence.
make cert creates a local signing identity once, so macOS keeps the
permissions across rebuilds.
To work on the interface with sample data and without the models:
cd frontend
VITE_DESIGN=1 npm run dev
Contributing
Issues and pull requests are welcome. AGENTS.md is the working
agreement for this repository: how the code is organised, what must not break,
and how changes are tested. Audio, model and threshold changes are decided by
measurements in exp/, described in experiments; the
roadmap lists what comes next.
License
MIT. See LICENSE.
Acknowledgements
Meeting Transcriber stands on these projects: whisper.cpp and llama.cpp (MIT), sherpa-onnx (Apache-2.0), audiotee (MIT), Wails (MIT), React, Motion, Lucide and Tailwind CSS, and the Geologica typeface (SIL Open Font License, notice).
The models it downloads keep their own licences: Whisper large-v3-turbo (MIT), Silero VAD (MIT), pyannote segmentation 3.0 (MIT), 3D-Speaker CAM++ (Apache-2.0), Parakeet TDT 0.6B v3 (CC BY 4.0), Gemma 4 E2B (Apache-2.0) and Qwen3 Embedding 0.6B (Apache-2.0). FFmpeg is downloaded as a separate program under its own licence.