MeetingDebrief

Your own AI server

Run summaries and chat on a server you already have

Overview

The built-in AI model takes several gigabytes of memory while it works. If you already run an AI server, locally or on another machine, MeetingDebrief can use it for summaries, titles, chat, and vocabulary suggestions. The built-in model remains the default.

What goes where

Your transcripts, notes, and chat questions go to the server you choose, so pick one you trust. Audio does not: recording and transcription always stay on your computer.

Supported servers

Any OpenAI-compatible chat server, including llama.cpp (llama-server), Ollama, LM Studio, and vLLM. Online AI providers such as OpenAI, Anthropic, and Gemini are not supported. If you run a server that doesn't work and would like to see it supported, get in touch.

Connect a server

  1. Open Settings → AI & Summaries and go to Where summaries run. On a new install, the setup wizard asks the same question under Summaries & chat; choosing a server there skips the summary model download.
  2. Enter the Server address.
  3. If the server needs a key, add it under Access token (optional).
  4. Click Test connection. MeetingDebrief detects the server type, its models, and its context size, and runs a short test answer.
  5. Click Use this server.

If the server isn't running yet, choose Save anyway, I'll start the server later after a failed test.

Advanced options

  • Model name: picked automatically unless you set one.
  • Context size (tokens): detected automatically where the server reports it. Ollama's default is small (4096), so raise it for long meetings.
  • Token header: for gateways that expect a header such as x-api-key.

Security

Tokens are stored in the macOS Keychain or Windows Credential Manager. They never go in the app database, logs, or feedback bundles. Plain HTTP works on your local network, and the app warns you before sending a token unencrypted to a public address. Only certificates your system trusts are accepted, so self-signed certificates don't work yet.

Long meetings

A meeting that doesn't fit the server's context window is split into parts automatically.

When the server is down

The job fails with a plain message such as Can't reach the AI server. MeetingDebrief doesn't switch to the built-in model and doesn't queue the job to retry later. Start the server, then generate the summary again.

Switching back

In Settings → AI & Summaries, use Change… to edit the connection or Remove server to return to the built-in model. While a server is active, you can delete local summary models to free disk space.

Tap image to zoom