Docs/Settings/AI

AI

This tab sets what every new conversation starts with — which model it uses and how hard that model thinks — plus the keys for the AI services you want to use, and voice dictation. If you want a particular thinking tool to use a different model, that's set on the Skills tab; here you're setting the defaults everything else falls back to.

Default model and effort

Two dropdowns pick the starting point for a fresh conversation. They're only defaults: each conversation has its own model and effort pickers that override them, and changing them here never disturbs a conversation you already have open.

Default model

Every model from every service you've set up, grouped by who makes it. Faster models answer quickly and cost less; more capable ones handle harder, more nuanced work.

Default reasoning effort

How long the model spends thinking before it answers, from Low up to Max — or leave it at the model's own default. More effort can mean better answers on tricky questions at the cost of a little speed. Not every model offers every level, and you're only offered the ones the chosen model actually supports.

Bring your own model

Minerva talks to four kinds of AI service. You only need to set up the ones you intend to use — a key for one is enough to get started, and you can add others later without disturbing anything.

Anthropic

The Claude models. Minerva is built around them and they're what the defaults point at.

OpenAI

GPT-5 and the o-series reasoning models.

Google Gemini

The Gemini 2.5 models.

Local / OpenAI-compatible

Anything that speaks the OpenAI protocol at an address you give it — Ollama or LM Studio on your own machine, a vLLM or llama.cpp server, or a gateway like OpenRouter or Together. See Running models locally below.

Keys

Each service gets its own section with its own key. A line above the field tells you whether one is saved, and says encrypted at rest when your operating system provides a secure store for it. For your privacy a saved key is never shown back to you and can't be copied out of the field.

Entering a key

Paste it into that service's field and click Done. If a key is already saved, typing a new one replaces it.

Clearing a key

A Clear saved key link appears once one is saved. Nothing is removed until you click Done, and Cancel clear lets you change your mind first.

Check connection

Confirms the service accepts the key before you rely on it — much easier than discovering the problem halfway through a conversation. It tests whatever the app would actually use, so it works on a key you've just typed as well as one already saved.

Environment variables

If you'd rather not store a key in the app, each service names the environment variable it will fall back to (ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY). A key saved here takes precedence over one in the environment.

You choose who sees your notes

Everything else in Minerva stays on your machine. A conversation is the exception — the parts of your notes the assistant looks at are sent to whichever service that conversation is set to use. Picking the model is therefore also picking who that is, which is the main reason the choice is per conversation and not buried in a preference.

Running models locally

The Local / OpenAI-compatible section takes a base URL instead of assuming one — http://localhost:11434/v1 for a default Ollama install, or wherever your server actually listens. Most local servers don't need a key; leave it blank if yours doesn't.

Minerva can't know what your endpoint serves, so you list the models yourself: type the model's id (the name the server knows it by, like llama3.1:70b) and, if you like, a friendlier label to show in the pickers. They then appear alongside the built-in models everywhere a model can be chosen. Remove one and it disappears from the pickers; conversations already using it keep their setting.

Voice dictation

Turn on Enable voice dictation to speak instead of type, and pick a voice model from the dropdown beside it. Dictation runs entirely on your own machine — your audio never leaves it. The first time you use it, the voice model downloads once and is then ready to use offline.