An MCP server is an external service that publishes a set of tools in the Model Context Protocol. Connect one and its tools appear next to your custom tools, ready to attach to an assistant, an agent or a workflow step. Open MCP servers in the workspace sidebar to manage them. Adding, syncing and removing servers is for workspace admins; every member can see them.
Only remote servers are supported (Streamable HTTP, a URL). Servers that run as a local process are not.
search on a server named notion becomes notion__search.With OAuth, Add and authorize takes you to the server to approve access, then back to the MCP servers page. A server whose authorization was never finished shows Authorize; one whose access has lapsed shows Reconnect, which keeps its tools and attachments.
Adding a server lists its tools and creates them in your workspace. Sync now repeats that: new tools are added, descriptions and schemas are refreshed, and tools the server no longer lists are removed (and detached from anything using them).
A synced tool's name, description and input schema belong to the server, so you can't edit them here. A sync only happens when you ask for it. Once a day the platform checks each server and, if its tools changed, shows a Server changed badge naming the differences, with Review and sync to apply them. It never changes your tools on its own.
The platform can't tell a read from a write, so new MCP tools start as require approval: a call pauses until a person approves it. Once you've looked at a tool and trust it, change its approval mode on the Tools page. MCP tools can also use a rule that asks only for some arguments, through the API.
Results from an MCP server are treated as outside content: an assistant is told not to follow instructions found in them. Output over 50,000 characters is cut off. If a workflow step can both read outside content (including MCP) and take an action without approval, the workflow form warns you.