Integrations
Integrations covers two things: the MCP servers that give your agents tools, and the project keys your own code uses to reach the platform.
MCPs
Connecting an MCP server here makes its tools available across the project. Agents opt in individually from the Connections step of their configuration.

The catalog covers common services, and connecting one asks for little more than a credential, since the endpoint, transport, and authentication method come with the connector. A few public connectors need no credential at all. Custom MCP connects anything else that speaks HTTP, where you supply the endpoint and authentication yourself and can add extra headers the server expects.
Either way the platform dials the server when you save and fetches its tool list, so a connection that comes back with a tool count is one that answered.
Working with a connection
Opening a connection shows how it is reached and what it offers: its endpoint and transport, the authentication in use, the MCP protocol version, and when it was last verified. Re-verify dials it again, which is the first thing to try when an agent stops getting results from a tool.

Every tool the server exposes is listed with its description, and Test runs one so you can confirm it behaves before an agent depends on it. Tools carry the connection's prefix, as in tavily:tavily_search, and that prefixed name is what you write when a policy or a routine names a tool. The prefix comes from the connection's name, so a name you would not want to type into every policy is worth changing before you attach it to anything.

Connected agents shows which agents reference the integration. Check it before disconnecting, since agents that reference the connection lose its tools.
Project keys
Project keys authenticate your own application against the platform through the SDK. Each key is a public and secret pair, with a note to tell them apart. They are not Router keys, which authenticate LLM calls and live on the Router page.

Last updated
Was this helpful?

