Search-MCP provides three well-scoped tools with solid naming conventions (all start with action verbs: search-*, fetch-*) and complete input schemas using Zod with type constraints. All tools have descriptions and all parameters have descriptions. However, the server has significant weaknesses: (1) output schemas are completely undocumented, handlers return generic `{content: [{type: 'text', text: JSON.stringify(...)}]}` with no guidance on what fields the LLM should expect in the JSON payload; (2) error handling is absent, no error messages, recovery guidance, or input validation visible; (3) parameter descriptions lack expected format/range details; (4) no pagination or result limits documented for search tools that could return many results; (5) tools are read-only but lack explicit security/idempotency markers.

• Document output schema for each tool. For search-web/search-github, provide a JSON Schema showing returned hits array structure (e.g., {hits: [{title: string, url: string, snippet: string}]}). For fetch-url, specify that the response is markdown-formatted text.
• Add error handling and recovery guidance to all handlers. Wrap external API calls in try-catch, validate inputs (e.g., URL format for fetch-url), and return structured error objects with actionable next steps: {error: 'rate_limited', message: 'DuckDuckGo rate limited. Retry in 60 seconds or use a shorter query.', recoverable: true}.
• Enhance parameter descriptions with concrete constraints. For search-web limit param: 'Max results to return (1-30, default 10). If omitted, returns up to 10 results.' For fetch-url: 'HTTP/HTTPS URL to fetch (max 10MB, timeout 30s, follows redirects up to 5 hops).'
• Document pagination behavior for search tools. Specify: 'Results are returned as a single page. To retrieve more, adjust the limit parameter. No pagination cursor is available, this tool returns at most N results per call.'
• Add tool annotation hints to indicate read-only safety. Use server.tool(..., {description: '...', readOnlyHint: true}, handler) to signal that all three tools are safe for retry.
• Consider adding a fourth tool (get-search-suggestions or similar) to offer query refinement if a search returns no results, enabling better multi-step planning.

Inferred effective spec: <=2025-11-25.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained