mcp_servers was set without mcp_toolset

The setup that looks complete but isn't

A request declares one or more MCP servers under mcp_servers, the servers themselves are correctly configured and reachable, and the request still fails — because declaring a server and giving Claude a way to actually call it are two separate steps, and only the first one happened here. The server connection exists; the tool interface that would let a model reach it does not.

Why declaring the server alone isn't enough

mcp_servers tells the API which servers exist and how to reach them, but it doesn't by itself put anything in front of the model to call — that's the job of including the MCP tool connector in the request's tool list. Skip that second half and you've built a bridge with nothing on either end of it: the server is live, but there's no tool definition in the request that Claude could invoke to reach it, so the request is rejected rather than silently proceeding with an unusable server.

Why this is easy to miss

It's an easy step to drop because it's genuinely a separate field from the one that feels like the "main" configuration — most people setting up their first MCP server focus entirely on getting the server itself defined correctly (command, arguments, environment) and treat the tool-list entry as an afterthought, if they remember it's a separate requirement at all. Once you've hit this once, the two-part shape is obvious; before that, it isn't documented anywhere near where the server config itself lives.

The fix

Add the MCP tool connector to the request's tools list alongside your mcp_servers declaration — the server config and the tool declaration are both required, and neither one substitutes for the other. If you're setting up your first server rather than debugging an existing one, see connecting an MCP server to Claude Code for the fuller walkthrough, which covers this pairing from the start rather than as a fix after the fact.