The MCP Registry server exposes 3 read-only tools with well-structured JSON schemas and reasonable descriptions. All tools follow verb_noun naming conventions (list, get, get). Schemas are complete with typed parameters, defaults, and constraints. However, descriptions are moderately detailed but lack LLM-optimized guidance on WHEN to use each tool vs. alternatives, and error handling is not documented. Output schemas are not explicitly defined in the tool descriptions, forcing LLMs to infer response structure. No destructive operations, which simplifies the safety model, but no tool annotations are present.

• Document output schemas for each tool. Example: 'Returns: {servers: [{name, description, latestVersion, updatedAt, ...}], nextCursor, total}'. This lets LLMs know what fields to expect and how to chain calls.
• Add WHEN guidance to tool descriptions. Example for list-servers: 'Use to discover available servers, filter by name/update time, or paginate results. For details on a specific server version, use get-server-version.'
• Clarify version parameter format. Document valid version formats (e.g. 'Semantic versioning (1.2.3) or the special value "latest"'). Include error cases: 'If version not found, returns a 404 error. Call list-servers with the server name to see available versions.'
• Add examples to descriptions of common use cases. E.g. 'Example: Find all servers updated in the last 30 days: call list-servers with updated_since set to 30 days ago.'
• Add readOnlyHint: true to all tool definitions to signal these are safe read-only operations per MCP spec 2026-07-28.
• Document pagination behavior explicitly. E.g. 'Returns up to 'limit' items (default 30, max 100). Use 'cursor' to fetch the next page. Iteration is complete when the response contains fewer items than requested.'