pi-mcp is a meta-tool that dynamically wraps other MCP servers' tools. Tool definitions are inferred at runtime rather than statically registered. The 4 tools shown have minimal descriptions (mcp_discover is adequate at ~100 chars, but mcp_<server>_<tool> and resource/prompt tools are generic placeholders). Input schemas are present but sparse: mcp_discover has a basic optional 'server' param; dynamic tools have empty schemas ({}); resource and prompt tools have minimal required params. No output schemas documented. No error handling guidance. Tool names use underscores but lack clarity about what 'mcp_' prefix means to an LLM. This is a proxy/gateway pattern, not a domain-specific tool suite, which limits scoring potential.

• Document output schemas for all tools. mcp_discover should return {tools: [{name, description, inputSchema}], resources: [{uri, name, mimeType}], prompts: [{name, arguments}]}. Resource and prompt tools should return {content: string, mimeType?: string}.
• For dynamic tools (mcp_<server>_<tool>), generate and cache input schemas from the upstream MCP server at discovery time, then expose them in the tool definition. Do not register tools with empty schemas.
• Add error handling guidance: 'If the tool fails, check that the MCP server is running. Retryable errors include timeout and connection refused. Non-retryable: invalid resource URI, tool not found.'
• Rename tools to clarify intent: 'discover_mcp_tools' (verb-first), 'call_mcp_tool', 'read_mcp_resource', 'get_mcp_prompt'. Avoid generic 'mcp_' prefix without context.
• Document parameter constraints in descriptions. E.g., 'server: optional string, filter by MCP server name (case-sensitive). If omitted, returns all servers.' Add examples of valid resource URIs and prompt names.

Inferred effective spec: 2026-07-28+.

• Relies on Roots (deprecated) - pass files/dirs via tool parameters, resource URIs, or configuration