mcp-lens defines 4 meta-tools with clear verb-noun naming (search_tools, describe_tool, call_tool, list_servers). All tools have descriptions (194 - 250 chars, within baseline 34 - 392 range) and explicit Zod schemas with typed parameters. Descriptions explain WHAT each tool does and WHEN to use it (e.g., 'Use this before call_tool to discover which server/tool to use'). However, output schemas are not documented, responses are wrapped as JSON text without formal schema declarations. Error handling is present (toErrorResult) but lacks recovery guidance (e.g., 'Call search_tools first to find valid names' is mentioned in describe_tool but not systematized). Parameter descriptions are adequate but could be more prescriptive about constraints (e.g., limit max=50 is enforced but not explained in the description text itself).

• Document output schemas formally. For search_tools, declare: 'Returns array of {name: string, server: string, tool: string, description: string, inputSchema: object, score: number}'. For describe_tool and call_tool, document the structure of returned tool metadata and call results.
• Add recovery guidance to error messages. E.g., describe_tool error: 'Unknown tool "X" on server "Y". Call search_tools(query="X") to find valid tool names.' call_tool error: 'Failed to call tool on server "Y": [reason]. Verify the server is connected via list_servers().'
• Expand parameter descriptions with constraint text. For search_tools limit: 'Max number of results to return (default 10, max 50).' For call_tool arguments: 'Arguments to pass to the downstream tool. Omitted required parameters will cause the downstream tool to fail, use describe_tool first to see required fields.'
• Add tool annotations via the MCP SDK. Mark search_tools, describe_tool, list_servers with readOnlyHint=true. Mark call_tool with destructiveHint=true (since it can invoke write operations on downstream servers). This helps LLMs reason about safety.
• Consider adding a validate_tool_call tool that dry-runs a call_tool invocation without side effects, returning validation errors before the LLM commits to the call. This prevents wasted calls and guides parameter correction.