Skip to content

Local models and AI providers

Granello does not have its own AI service. In Settings → AI models you choose the provider and the model, and you can use a different one for the AI notes and for the chat.

Provider type Used for
Anthropic Claude, the default provider
OpenAI-compatible LM Studio (http://localhost:1234/v1), Ollama (http://localhost:11434/v1), OpenAI, OpenRouter and any server that exposes /chat/completions
  • The list of models is read from the server. Test checks that the provider answers.
  • Each provider has its own API key in the Keychain. Local servers usually do not need one.

Add a local provider (a server on localhost) and select it for both AI notes and Chat. From then on, no text leaves your Mac.

To make sure of it for one meeting whatever the settings say, set that meeting to hardened: it only ever uses a provider on localhost.

An hour of meeting is tens of thousands of tokens. Local servers silently cut what does not fit, and some start with a small context. Enter it in the provider’s Context (tokens) field: Granello then estimates the size of each request and stops with a clear error, instead of writing notes from half a transcript.

The reasoning that some local models print before answering is removed from the notes.

Speech is transcribed on your Mac by whisper.cpp, whatever AI provider you use. Sending audio to an online transcription service is a separate option, off by default: see how privacy works.