MCP servers and tools
MCP servers give agents extra tools — reading from a system, calling an API, querying a database — without changing the agent itself.
Adding a server
Section titled “Adding a server”Settings → Tools manages MCP servers. Four transports are supported:
- stdio — a local command the app launches
- SSE — server-sent events over HTTP
- HTTP and streamable HTTP — remote servers
Add a server manually, or paste an existing MCP JSON configuration to import one in a single step. Remote servers requiring OAuth trigger their sign-in flow when you test the connection.
Choosing which tools are exposed
Section titled “Choosing which tools are exposed”Each server lists the tools it provides, and tools can be enabled or disabled individually. This matters for two reasons: a smaller tool list is easier for a model to use well, and it keeps capabilities you do not want out of reach.
Per-assistant selection
Section titled “Per-assistant selection”Which MCP servers an assistant uses is part of that assistant’s configuration, alongside its skills — so a “research” assistant and a “coding” assistant can carry entirely different toolsets.
Like other assistant defaults, MCP selection can be auto (follow your last used set) or fixed (pinned to this assistant). See Assistants, agents, and models.
In conversation
Section titled “In conversation”The conversation shows which MCP servers loaded for the current session, so you can confirm an agent actually has the tools you expect before relying on them.