Setting up the AI tools

LymeScribe's History actions — Summarize, Meeting Minutes, and Translate — don't come with a built-in AI. You point them at an inference engine you choose: a model running on your own machine, or a cloud provider's API. Here's how to set one up.

Last updated 2026-08-31

Where this lives

Open Settings and go to the AI pane (section header "Inference Engines"). This is where you add, edit, and remove engines, and pick which one runs by default when you use a History AI action. It's empty until you add your first engine.

Two paths: local or cloud

Local (fully on-device)

Run a model on your own computer with LM Studio or another OpenAI-compatible local server (Ollama works too). Nothing about the transcript ever leaves your machine — not even to LymeScribe's own servers, since there aren't any in this path.

Cloud provider

Point LymeScribe at a hosted API — OpenAI, OpenRouter, or any other service that speaks the OpenAI chat-completions format. Faster to set up, no local hardware needed, but the transcript text is sent to that provider under their terms.

The privacy line: nothing leaves your machine unless you point a profile at a cloud endpoint. A local engine (LM Studio, Ollama) keeps summaries, minutes, and translations entirely on-device, same as the transcription itself.

1Add an engine

In the AI settings pane, click Add Engine. There are shortcuts for the common cases — LM Studio (local), OpenAI, OpenRouter — or choose Custom to fill in any OpenAI-compatible server by hand.

Picking a preset prefills the base URL (LM Studio defaults to http://localhost:1234/v1). Custom starts blank.
Give it a name, e.g. "Mac Studio (LM Studio)" or "OpenAI".

2Fill in the engine's details

Base URL: The address of an OpenAI-compatible server, up to and including /v1 — e.g. http://localhost:1234/v1 for a local LM Studio instance, or https://api.openai.com/v1 for OpenAI.
API Key: Required for most cloud providers, usually unnecessary for local engines. Stored in your machine's secure credential store (Keychain on Mac), never in plain settings files.
Model: The model identifier the engine expects, e.g. qwen2.5-32b-instruct or gpt-4o. If the engine is reachable, LymeScribe can pull its available models for you to pick from instead of typing one.
Context Budget: How many tokens the engine can accept in one request. When you pick a model from a local engine that reports its own context window, this fills in automatically — otherwise it defaults to a conservative 32,000 and you can raise or lower it.

Use Test Connection before saving — it confirms the engine is reachable and, if so, reports how many models it found.

3Set a default

Once you've added at least one engine, pick a Default Engine in the same pane. That's the one History AI actions run against unless you choose a different one in the action's own pre-flight step.

Long transcripts and chunking

A short recording fits inside an engine's context budget in one request. A long one — a two-hour meeting, say — may not. When it doesn't, LymeScribe splits it into ordered chunks that each fit the budget, running Summarize and Meeting Minutes as a map-reduce (each chunk summarized, then combined) and Translate chunk-by-chunk in sequence. You don't have to do anything for this to work; it's automatic based on the context budget you set for the engine. A larger context budget means fewer, larger chunks and a faster result — one more reason to raise it when your local model can handle it.

Running an action

From any transcript in History, choose Summarize, Meeting Minutes, or Translate. The result streams into a sheet as it's generated and is saved onto that transcript, so you can come back to it later without re-running the action.

Ready to try it? Download LymeScribe — the AI actions are part of the app; you bring your own engine. For prompt ideas once you're set up, see prompt recipes for transcripts. Questions? Email us — a real person answers.