Three tools with clear verb-noun naming (find_mcp, read_mcp, search_mcp) and solid descriptions (100-150 chars each). All tools declare READ_ONLY annotations. Input schemas are present with Zod types and descriptions. However, output schemas are not documented in the visible code, the handler signatures return `Promise<unknown>`, leaving LLMs unable to plan downstream calls. The task_context parameter is well-designed for telemetry but adds friction to every call. Error handling and recovery guidance are absent from descriptions.

• Document output schemas for all three tools. E.g., find_mcp returns {title: string, summary: string, sections: {heading: string, links: {title, url}[]}}. This enables LLMs to extract and chain results.
• Add error recovery guidance to descriptions: 'If no llms.txt is found, try searching the domain root or checking if the site publishes llms.txt at all.'
• Clarify search_mcp query semantics: 'Queries are case-insensitive substring matches against section headings and link titles. Returns up to 20 matches; use read_mcp for full content.'
• Consider making task_context optional with a default value (e.g., 'general discovery') to reduce required parameters and improve usability.
• Add pagination to read_mcp: accept optional limit and offset parameters, return total_sections count, and document max response size (e.g., 2 MB cap visible in code).

Inferred effective spec: 2025-06-18+.

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