Skip to main content

Connectors

A connector is an MCP server the agent can call tools through — GitHub, a database, a browser, your files. It is how the agent reaches something that is not a file on this machine.

Adding one​

Abilities → Connectors lists what is configured and what can be added. Adding from the catalogue takes one click; adding your own takes a name and either a command or a URL:

Local commanda program on this machine, started by OpenCLI
HTTP servera server reachable over HTTP

Keys​

Most connectors need one. Name the environment variables the server expects — names only, separated by spaces — and set the values afterwards.

Values are kept with your other keys, not in the connector's configuration, and are never shown again once saved. That is deliberate: the configuration file gets shared and pasted into issues, and a key in it goes with it.

Testing​

Test connections starts every configured server and waits for a handshake. It is a button rather than something that happens when the panel opens, because starting them takes seconds — measured at 2.2 against one that fails to authenticate.

The result appears on each row, with the tools that server offers.

When a change takes effect​

Servers start with a chat. A connector added now applies to the next conversation you open, not the one in front of you.

What a connector is not​

It is not a capability the model gains. A model that will not call tools does not start calling them because a connector was added — see Limits.

Departments​

The chips above the list filter by department, so a newly created Finance department can be asked what should this be connected to rather than what is Playwright.

A connector with no department tag shows under every chip: something nobody has categorised should not be hidden from everybody.