Local Edition does not decide which AI model to use. You do, in Settings, AI engine. Only one engine is active at a time, and the active engine is shown at the top of every chat so you always know which model produced a map.
The three kinds of engine
- Bundled local model. Ships with the app and runs entirely on your computer. No account, no key, no network. It is the most private option and the slowest, and it produces simpler maps than the large cloud models. It supports tool calling, so the Agent works with it.
- Local endpoint. A model server on your own machine or network that speaks the OpenAI-compatible API, for example a model you run yourself. Enter the server URL and the model name.
- Bring your own cloud model. A model from a cloud provider, billed to your own account with that provider: OpenAI, Anthropic, or any provider with an OpenAI-compatible endpoint (Google, xAI and Mistral among them). Enter your API key once; it is stored in your operating system's credential store, never in the database.
Add an engine
- Open Settings, AI engine and choose Add engine.
- Pick the provider. For OpenAI and Anthropic the endpoint is filled in for you. For an OpenAI-compatible provider, paste the base URL from that provider's documentation.
- Enter the model name exactly as the provider spells it (see the recommendations below for the names to paste).
- Paste your API key and choose Test connection. A green result means the key and model name work.
- Set the options:
- Supports tool calling. Leave on unless the provider says the model cannot call tools. The Agent needs this; the chat does not.
- Supports vision. Turn on if the model accepts images, so you can attach screenshots and diagrams.
- Options per request. How many alternative maps a generation returns, from 1 to 3. More options cost more.
- Concurrency. Sequential is safer for rate-limited accounts; parallel is faster.
- Choose Activate. The engine name appears in the chat header.
You can keep several engines and switch between them at any time. Switching does not change existing maps.
Which model should I use?
We cannot recommend one model for everyone. Whether a model meets your needs depends on the processes you describe, how complex they are, whether you attach documents or images, how much you use the Agent, and what you are willing to pay per generation. Models also improve every few months, so a recommendation goes stale quickly. Try the newest models from your provider and keep the one that gives you the maps you want at a price you accept.
That said, here is our recommendation as of 17 September 2026, based on our own testing:
| Rank | Model | Provider and model name to paste | Chat generation | Agent | Cost |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.8 | Anthropic, claude-opus-4-8 | Best | Best | Highest |
| 2 | GPT-5.6 Sol | OpenAI, gpt-5.6-sol | Great | Good | Moderate |
| 3 | GPT-5.6 Luna | OpenAI, gpt-5.6-luna | Great | Not recommended | Lowest |
- Claude Opus 4.8 produces the most complete maps from short descriptions and is the strongest at multi-step Agent work such as building a linked hierarchy of maps. It is also the most expensive per request.
- GPT-5.6 Sol is a strong all-rounder: excellent maps in chat, good Agent results, and noticeably cheaper than Opus.
- GPT-5.6 Luna is the value choice for chat generation. It is fast and cheap and produces good maps, but it does not follow the Agent's multi-step instructions reliably, so keep it for chat.
The bundled local model is not in the table because it is free and private but not comparable on quality. It is a good way to try the product without any account, and a reasonable choice for straightforward processes.
If a model does not behave
- "The active engine can't call tools." The engine's tool-calling option is off, or the model really cannot call tools. Turn the option on, or pick another engine for Agent work.
- Generation stops with no map. Some providers stall on long requests. Try again, reduce the options per request to 1, or switch to a model with a larger context window.
- Slow first response on the bundled model. The model loads into memory on first use after a restart. Later requests are faster.