AI providers
Which model each organization uses, and how to run your own.
Which model answers
For an organization, the API uses the first of:
- The organization's own provider, set by its owner under Control Center › AI Providers.
- The global provider, which a super user sets for every organization without one.
- The fallback,
ANTHROPIC_API_KEYin the server's environment.
With none of them, the organization's AI features are off: mail still syncs, it just isn't sorted or summarized.
Provider kinds
| Kind | For |
|---|---|
openai | Any endpoint that speaks the OpenAI chat format, at any base URL. |
anthropic | Anthropic's API, or anything compatible with it. |
self_hosted | Your own server in OpenAI format, such as Ollama, vLLM or LM Studio. The key is optional. |
The form lists the models the endpoint offers. A saved key is only reused for the same kind and base URL, so pointing a provider elsewhere asks for the key again. Keys are encrypted with APP_SECRET.
A super user can move a provider between an organization and global.
A local model
To use Ollama on the same machine or network, set ALLOW_PRIVATE_HOSTS=true on the server (it lets base URLs reach private addresses and use http), then add a self_hosted provider with a base URL such as http://ollama:11434/v1.
ALLOW_PRIVATE_HOSTS also lets users point IMAP servers and MCP servers at your private network. Keep it off on a server where people you don't trust can sign up.