Run an AI agent on your Mac with no cloud at all

Canoryn works with the model you choose: a frontier model with your own API key, your ChatGPT plan, or a model running on your own Mac. This post is about the last option — the one where nothing leaves your machine.

Why run locally

• Privacy. Your prompts, your code and the model's output never touch a server.
• Cost. After the download, a local model is free to run, as often as you like.
• Offline. On a plane, on a train, on bad hotel Wi-Fi — it still works.

The trade-off is capability and speed. Local models are smaller than frontier models, so they're best at focused jobs: summarising a page, drafting a commit message, answering questions about a file, running a workflow step over and over.

Setting it up

1. Install Ollama and start it. It runs quietly in the menu bar and serves models on your Mac.
2. Pull a model. In Terminal, for example: ollama pull qwen3:8b. Pick a size your Mac's memory can hold comfortably — roughly, the model's download size plus a few GB.
3. Add Ollama in Canoryn as a provider. Canoryn finds the models you've pulled.
4. Choose it for a chat session, or for an AI step on a workflow canvas.

That's it. The agent loop is the same one you'd use with a cloud model: it reads files, runs commands and asks before anything risky — just with the model on your Mac.

Mixing local and cloud

You don't have to pick one for everything. A common setup is a frontier model for the hard, one-off job in a coding session, and a local model for the workflow that runs every morning and summarises the same five pages. Each session and each AI step picks its own model.

Tips

• Give it a smaller job. Local models do better with one clear task than a vague, sprawling one.
• Watch the context size. Smaller models have shorter memories; long documents may need to be split.
• Keep one fast model around for quick steps, and a bigger one for when quality matters more than speed.

Download Canoryn and try it with a model you already have.