Manage MCP servers from CHAT: a cheap, always-in-context CATALOG of the MCP tool servers available to this agent (name, short description, active/off/disabled), a read-only `mcp_catalog` tool, and a `/mcp` slash command to toggle a server's tools on/off per session — live, persistent, no restart.
The mcp_catalog tool has a clear, well-written description (read-only catalog of MCP servers). However, the tool definition lacks a visible input schema in the provided source code. The function signature and implementation are present, but no explicit JSON Schema with parameter types and descriptions is shown. Without verifiable schema documentation, the tool cannot score higher. The description itself is good (~100 chars), but absence of input schema documentation is a critical gap per the rubric's hard rules.
Read-only catalog of MCP servers available to this agent, showing name, short description, and active/off/disabled state
No input schema visible for mcp_catalog tool. The tool definition in __init__.py shows the function build_catalog() but no explicit JSON Schema registration with parameter types and descriptions.
Tool registration mechanism not visible in source. The code references 'tools.registry' and 'gateway.run' but does not show how mcp_catalog is registered as an MCP tool with schema. Tool definitions must be explicitly registered with names, descriptions, and input/output schemas.
Output schema not documented. The build_catalog() function returns List[dict] with fields 'server', 'tools', 'state', 'description', but no formal output schema is provided to the LLM. LLMs need documented return types to plan downstream operations.
No error handling guidance. The code has fail-open patterns (return None/empty on exception) but does not document error cases or recovery steps for the LLM. If the catalog fails to load, the agent has no guidance on what to do next.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 33 | 2026-07-28+ | v2 |
Tool does not accept parameters. mcp_catalog appears to be a zero-parameter tool (read-only catalog), but this is not explicitly documented. If the tool accepts optional filters (e.g., state='active', server_name pattern), those parameters must be declared with types and descriptions.