Set up a provider

A provider is the service that runs the AI model. Providers are configured in the AI Settings dialog. Open it with AI › AI Settings… or the gear in the chat header.

The dialog has one tab per provider and the tabs General, MCP and Isolation. Changes apply after Save.

The Claude tab with backend mode, default model, worker model and the limits for a turn.

Choose the default provider

1. In AI Settings, open the General tab.
2. Choose the Default Provider.
3. Click Save.

AI › Default Provider changes the same setting. New sessions use the default provider. A running session keeps its provider. To use another provider, choose it in the provider list of the chat header. That starts a new session.

Claude (CLI)

This provider runs the claude program (Claude Code) installed on your machine, with the login of that program.

Prerequisite: claude is installed and you have logged in once. The slash command /login runs the login from the chat.

On Windows, the subscription mode is not available. The chat uses claude -p.

Anthropic (API)

This provider calls the Anthropic Messages API directly. claude does not need to be installed.

1. In AI Settings, open the Anthropic tab.
2. Choose the Authentication:
   • API key (Console billing): your own key from the Anthropic Console, billed per token. Enter the key under API Key and click Save next to the field.
   • Subscription (billed as extra-usage): reuses the login token of claude. No key is needed. Turns are billed against the extra usage of your Claude plan.
3. Click Test & Load Models. The model list is filled with the models your account offers.
4. Choose the Default Anthropic Model.
5. Click Save.

With Subscription, the tab shows whether extra usage is enabled on your account and how much of it is used. If extra usage is disabled, turns stop with an error once the regular quota is used up.

"Subscription" means different things in the two Claude providers:

OpenAI and OpenCode

OpenAI (API) uses the OpenAI API. OpenCode (Zen gateway) uses the OpenCode Zen gateway, which offers models of several vendors behind one key.

1. Open the OpenAI or OpenCode tab.
2. Enter the API Key and click Save next to the field.
3. Click Test & Load Models.
4. Choose the default model.
5. Click Save.

Both APIs accept at most 128 tools per request. When the tools of Tracemoor , the reference projects and external MCP servers exceed that number, Tracemoor drops the tools at the end of its priority order. The file tools, external MCP servers, the core tools of the active project and the xref_* tools come first, so the less used tools of Tracemoor are dropped. Claude and Anthropic have no such limit.

Ollama

Ollama (local + cloud) uses an Ollama server.

1. Open the Ollama tab.
2. Check the Base URL. The default http://localhost:11434 fits a standard installation.
3. Optional: enter a Cloud API Key (ollama.com) to also list the hosted models of ollama.com, marked "(cloud)". These run on ollama.com, not on your machine.
4. Click Test & Load Models and choose the Default Ollama Model.
5. Click Save.

Only models that support tool calls are listed. Models without tool calls cannot use the index.

Custom providers

A custom provider is any server with the OpenAI Chat Completions API, for example mlx-openai-server, vLLM or LM Studio on your own machine.

1. Open the Custom tab and click + Add Custom Provider.

2. Fill in the fields:

   Field	Example
   Title	Gemma 4 26B local
   Model	mlx-community/gemma-4-26b-a4b-it-4bit
   API Base URL	http://localhost:8080/v1
   API Key	not-needed for most local servers

3. Click Save.

Each custom provider appears in the provider list of the chat header. Its key is stored in the settings file, not in the keychain.

Where keys are stored

API keys for Anthropic, OpenAI, OpenCode and the Ollama cloud are stored in the keychain of the operating system, never in ai-settings.json. Clear next to a key field removes the key.

Development builds (npx tauri dev) store keys in a protected file instead, because every rebuild has a new code signature and macOS would refuse it access to the keychain entries of the previous build.

Change model and effort during a session

• Model: the model list in the input bar, AI › Change Model, or /model <name>.
• Effort: the dots in the input bar, AI › Effort Level, or /effort <level>.

The provider of a session cannot change.

Other settings on the General tab