Configuration

Configuration

Agent Jo can run on a remote API, a local model, or both together — with rules-based routing sending light work to the local model and hard work to the strong one.

Model engines

Engine	What it is	Cost	Leaves the machine?
Anthropic (Claude)	Hosted frontier model	Paid per use	Yes — to Anthropic only
OpenAI-compatible	Any endpoint (self-hosted or provider)	Varies	Depends on endpoint
Ollama	Local models on your own GPU	Free	No

Fully offline

Point Agent Jo at Ollama and nothing leaves the machine at all:

1. Install from ollama.com
2. ollama pull qwen3
3. In the app, select Ollama as the engine

API keys

The first run asks for a key (if you chose Anthropic). It's sealed on this machine — encrypted at rest and never sent anywhere except the engine provider. See Security-and-Privacy for the encryption model.

Routing rules

Light work (classification, extraction, reformatting) can be routed to the local model; hard work to the strong one. The app tells you which rule fired for any given request.

Fine-tuning

Agent Jo can fine-tune a local model for a domain — and refuses to register the result unless it beats the model it started from. A fine-tune that scores worse than its base is reported as worse.