Models and providers
Bring your own model, and pick a different one for each kind of work.
motita does not come with a model. It talks to the one you already use, and the loop around it is the same whichever answers.
Providers
| Provider | How you connect |
|---|---|
| OpenAI, and any OpenAI-compatible host | API key |
| ChatGPT (Codex) | Sign in with your ChatGPT plan |
| GitHub Copilot | GitHub device code |
| Anthropic | API key |
Claude subscription (claude-code) | claude auth login in Anthropic's own client |
| Google Gemini | API key or your own Google login |
| Qwen | DashScope API key |
| Ollama | Nothing on your machine; an API key for Ollama Cloud |
An API key always wins over a stored login. Keys you paste go to a separate owner-only file beside the configuration (~/.motita/motita.env), so the configuration itself can be committed and shared. A key is never sent back to the browser.
From the browser
Settings → Providers lists every provider, where its key comes from, a field for a key and endpoint, and a button that tests the connection by listing the provider's models. Copilot, Codex and Gemini can sign in with an account from the same page (Qwen takes a DashScope key: its free account login ended on 2026-04-15).
A model for each kind of work
The Models tab sets the default model and, under it, one for each kind of work. A kind left on Same as default follows the session.
| Kind | Used for |
|---|---|
plan | Plan-mode turns |
code | Task mode, and background agents that write |
review | Background agents with the review role |
research | Background agents with the research role |
summary | Titles, compaction summaries, auto mode's guess |
A strong model for planning and a cheap one for summaries is a common split. If a kind's provider has no key, it falls back to the session's model and the log says so.
Check that a model did not get worse
Under the models, Quality check runs five small fixed tasks (create a file, fix a typo, count lines, rename a file, sort lines) on the model that does the coding. Each one runs in a temporary folder of its own, and a command that looks at the real result decides whether it passed. The page compares the run with the last one: fewer tasks passing, clearly more steps or clearly more cost is shown as worse.
It uses part of the model's quota, so it only runs when you ask and confirm. After you change the model, or the instructions the agent works with, the page tells you the last check is out of date.