MCP servers

Servers rook itself, and the agents under test, may reach. Flags mirror claude mcp.

rook mcp list --json
rook mcp get <name> --json
rook mcp add <name> -- <command> [args...]
rook mcp add <name> -t http --url https://host/mcp -H 'Authorization: Bearer ${TOKEN}'
rook mcp add <name> -s project -e KEY=VALUE -- <command>
rook mcp remove <name> -s project
rook mcp enable <name>
rook mcp disable <name>
rook mcp approve <name> --origin discovered

Scopes: user, project, local (default). project scope is .testmuai/rook/mcp.json, a file rook itself writes — a different thing from a server explore finds already declared in the target. approve trusts a project or discovered server after rook has read what it runs; those are the only two origins that wait for it. A user or local server is used as declared, with no approval step.

When explore finds .mcp.json in the target, those servers appear with origin discovered and need approve before a scenario can exercise them. Values stay ${VAR} references in get output.

Since 0.1.14 an http server connects over Streamable HTTP. A project or discovered one waits for approve like a stdio server, and a changed host needs approval again. get and approve print sends the value of VAR to <host> for each variable an http server receives, and values of 8 or more characters expanded from ${VAR} are removed from what it sends back. Each request has a 20 s deadline and an 8 MiB limit, and redirects are refused. sse and ws servers stay unsupported-transport. On every transport, a tool list without tools, a result without content or a text block with no text is a malformed response, not an empty success.