Connecting an MCP server
A URL and a token. The server describes the rest — except how dangerous it is.
An MCP (Model Context Protocol) server describes its own capabilities, so connecting one imports every tool it exposes along with the argument schemas it declares. It is the least configuration of any connection: a URL, and a token if the server wants one.
What you need
The address of an MCP server that speaks Streamable HTTP, and whatever it needs to let you in.
- Server URL — the MCP endpoint, usually ending in
/mcp. - Bearer token — if the server is authenticated. Leave it blank if not.
- Additional headers — anything else the server expects, such as a tenant or workspace id.
Streamable HTTP transport only. A local stdio server cannot be connected — that would mean running your process on our infrastructure.
How risk gets decided
Caterfli infers it, and errs towards caution. MCP describes what a tool takes and what it returns, but not how dangerous it is.
So the platform reads an operation type from each tool's name and any hints the server provides. An unrecognised verb is treated as an action rather than a read. An action inherits a higher risk level, which means an approval rule can catch it.
If a tool is marked riskier than it is
Expect this after connecting a server, and correct it on the tools page. A tool guessed wrong in the safe direction stops for a person. Guessed wrong the other way, it would not.
Sources and further reading
- Model Context Protocol specification: Transports (Streamable HTTP) — modelcontextprotocol.io
- Model Context Protocol specification: Tools — modelcontextprotocol.io