Your AI, briefed on every meeting.

“What did Sam say about the launch date?”

trekit-mcp is a Model Context Protocol server for your recordings. Add it to Claude Code, Claude Desktop or any other app that runs local MCP servers, and your AI can list your sessions, search them, read whole transcripts and pull out everything one person said. The apps have the same search built in, for when you only want the line.

# coming soon: the PyPI release is not out yet
$ claude mcp add trekit -- uvx trekit-mcp

Sample data. The MCP tools and search exist; the conversation is invented. uvx trekit-mcp is coming soon, once the package is on PyPI. The command leaves out the server address and token, and Claude Desktop is set up through its mcpServers config.

Read-only

Four tools: list_sessions, search_sessions, get_transcript and what_did_person_say. They read your sessions and change nothing.

Tags vs guesses

Each line includes a speaker_source field, user when you tagged the speaker and auto when the voice matcher guessed, so the AI can treat a guess as a guess.

Import

Bring in an audio or video file and it is transcribed, matched to speakers and available to your AI like any live session.

Start recording without opening the app.

Press the Action button or the Control Center control and the meeting is recording. The Lock Screen and Dynamic Island show the timer with Pause and Resume, so you never unlock to check on it. On a Mac, the recorder lives in the menu bar.

Thirty-second chunks. Filled: uploaded. Outlined: queued on the device. Red: recording now.

Lossy network

Audio is saved in thirty-second chunks and queued on the device. With no route to your server, chunks stay in the queue and upload once you are back on your Wi-Fi. A damaged chunk costs thirty seconds of transcript, never the meeting.

Phone calls

A phone call pauses the session. If the microphone cannot restart after the call, a Resume button appears on the Lock Screen.

Tag a voice, and Trekit can suggest it in later sessions.

During setup you read a short passage aloud, about 25 seconds, so your own lines are labelled from the first recording. Tag anyone else in a transcript and Trekit builds a voiceprint for them. As you tag more clips of the same voice, it starts suggesting that person in later sessions, and re-checks older sessions for the same voice.

Sample lines. An automatic label is marked auto; a near match shows up as a question you can accept or reject.

Guesses are guesses

Automatic labels carry the auto mark, and a near match is phrased as a question, never stated as fact.

Automatic titles

Until you rename a session, its title is the start time and the people who talked most, like 14:05 - Sam, Priya.

Kept on a Mac you own.

The server runs at home on Apple Silicon. Audio, transcripts and voiceprints sit in one SQLite file and a folder of audio on that Mac. Run it this way and there is no Trekit account to make and no Trekit service to send them to.

Away from home, chunks stay queued until you are back on your Wi-Fi, or go over your own Tailscale network if you set one up.

Pairing

A QR code holds the server's address, a device token and the fingerprint of its certificate. The apps reject any other certificate, and the server advertises nothing on the network.

Unnamed voices

Voices you have not named stay inside their own session. Nobody gets a profile until you name them.

No sign-ups

Every model is ungated, so there is no Hugging Face account, API key or Trekit login between you and a working server.

Archive

hindsight export writes the database and every audio file to one checksummed archive. hindsight import restores it on another machine.

What does leave your network

Model weights, once

A few GB from Hugging Face the first time the server starts.

Place names

If you allow Location, the phone sends it to Apple's geocoder to get a place name for the session. Recording works without it.

What your AI reads

When Claude or another AI tool reads a transcript through MCP, those lines go to whoever runs the model, as anything you paste into it would.

Nothing else

Uploads from outside your home go only through a Tailscale network you set up yourself.

Not built yet

Server in the Mac app Planned

One Mac app, distributed outside the App Store, records and can also run the server in the background and pair your phone by QR code. The server part stays optional. Today it starts from a terminal.

A cloud plan Planned

If you would rather not keep a Mac running, upgrade inside the app and a hosted server does the processing. The apps use one documented API either way, so they stay the same.