Providers
Connect an LLM provider so your agents have a model to think with.
A provider connection links Compass to a large-language-model service. It's the "brain" an agent uses to reason, and the model an Agent node runs in a workflow. Every agent needs one.
Available providers
Anthropic
Claude models via the Anthropic API.
OpenAI
GPT models via the OpenAI API (and OpenAI-compatible endpoints).
LM Studio
Local models served through LM Studio's OpenAI-compatible API.
OpenAI and Anthropic are hosted services you reach with an API key. LM Studio runs models on your own machine or network, which is useful for local development and keeping data in-house.
Connecting a provider
Start a new provider connection
Under Forge → Connections, choose a provider and click New connection.
Name it
Give it a clear name, like "Anthropic — Production", so it's easy to pick later.
Add credentials
Enter the provider's API key. For LM Studio, point the connection at your local server's base URL instead.
Choose a default model
Compass fetches the models your account can access; pick the one this connection should use by default. (Agents can still be set to a specific model.)
(Optional) Route through a gateway
If you use a gateway, select it here so this provider's calls are proxied through it.
Save
Compass verifies the key and shows the connection as Connected.
Choosing a model
The default model you set on the connection is what agents use unless they specify otherwise. When you build an agent, you select a provider connection as its model; you can keep the connection's default or choose another model the provider offers. See Building an agent.
Tip
Keep separate connections for separate purposes — for example one for a fast, inexpensive model used in testing and another for your production model. Agents reference the connection, so switching is a one-field change.
Rotating keys
To rotate an API key, edit the connection and enter the new key. Because every agent points at the connection rather than at a raw key, the change takes effect everywhere at once — no need to touch individual agents or workflows.