Connect model providers

Every model comes from Settings → Model providers. Pick a preset or enter a custom OpenAI-compatible endpoint; each provider has its own API key, model list, and reasoning switches.

Add a provider

1. Open Model providers

   Press Ctrl+, (⌘+, on macOS) to open Settings, then click Model providers.

2. Click Add provider and pick a source

   The menu lists presets; the last item is “Import from this computer…”. Presets fill in the name, protocol, and official base URL.

3. Fill in the connection

   Name, API protocol, base URL, and API key. The official URL can be left blank; custom URLs must start with http:// or https://.

4. Add models

   Use Fetch models to pick from the server, or type a model id below and click Add. The Reasoning switch next to each model decides whether sessions can offer thinking levels; models known to the catalog show a read-only badge instead.

5. Save and enable

   Check the Enable switch at the bottom and save. The card switch can disable or re-enable a provider at any time without losing its key.

Test, edit, and remove

Test connection reports latency and how many models the server returned; it is disabled without an API key. Edit opens the same form and shows only the last four characters of a saved key — leave it blank to keep it, or use Clear key.

Deleting warns that roles referencing the provider must choose a new model. After deletion, those role models show as unavailable until you pick a replacement under Settings → Roles.

Import from this computer

“Import from this computer…” reads providers already configured in Claude Code, Codex, and CC Switch. Configs without an API key or a custom URL are skipped. Pick sources, confirm the providers to import, and choose the models to add for each.

How keys are stored

Keys

Plaintext API keys stay in the main process; the UI shows a mask. They are stored in local config files, not an encrypted vault. Don’t share the app data folder, config exports, or logs that contain keys. After you delete a provider, roles that used it need a new model.