A provider is the service that runs the model. Desktop stores the provider choice separately from your workspaces and sessions, so changing projects does not require entering the same credential again.
Provider readiness model
Desktop treats configured, reachable, compatible, and qualified as separate states. The versioned provider registry records model identity and declared capabilities; the canonical Kestrel profile defines which capabilities the task path requires.
Hosted providers
OpenRouter, OpenAI, and Anthropic require an API key. Paste the key into Desktop settings, enter the exact model ID you want to use, and save the provider. Desktop does not write the key into project files, prompts, or transcripts.
Local providers
Ollama and LM Studio expose models from your computer through a local OpenAI-compatible endpoint. Start the provider first, load the model you want, and then enter that model's exact ID in Desktop. A key is normally unnecessary.
Verify the connection
Create a short test session and ask for a one-sentence response. If Local Core is healthy but the request fails, check the provider status, endpoint, credential, and model name before restarting Desktop.
Choose a compatible model
Use a registry identity whose capability checks satisfy the active profile. Tool calling, structured inputs, context limits, and reasoning formats are explicit provider capabilities rather than assumptions derived from the model name.
Change or remove a provider
Changing the default affects future turns; prior runs retain their effective provider/model evidence. Remove a credential only through the supported settings path and confirm no active run depends on it.