Single-tool server with solid naming and schema definition. The 'read_website' tool follows verb_noun convention and has a complete JSON Schema with proper type constraints. Description is adequate but could be more specific about LLM-relevant behavior (when to use, dependencies, output format). Parameters are well-typed with min/max bounds. However, output schema is completely undocumented, LLMs cannot see what fields will be returned, violating the 'document output schema' critical check. No error handling guidance; no recovery paths documented. Tool annotations present but underutilized.

• Document the complete output schema for read_website. Include all returned fields (markdown, error, url, title, etc.), their types, and when each is populated. This enables LLMs to compose with downstream tools.
• Expand tool description with LLM-relevant guidance: 'Use this tool to extract content from documentation sites, blogs, and publicly accessible web pages. Best for reading structured content; less suitable for heavily JavaScript-rendered sites. Returns markdown + metadata. Call search_google or similar first if you only have a domain name, not a full URL.'
• Add explicit error handling hints: 'If the URL returns 401/403, try with cookiesFile parameter. If markdown is empty but error is null, the page likely has no readable content. If timeout occurs, retry with default timeout or contact support.'
• Add input validation documentation: 'cookiesFile must be a path to a valid Netscape format cookie file. The path is resolved relative to the server's working directory; no traversal above the server root is allowed.'
• Include 'pages' parameter guidance in description: 'pages > 1 enables crawling; crawls up to N pages starting from the URL. Crawling is depth-limited to 1 level to avoid runaway traversal. Set pages=1 (default) for single-page extraction.'
• Document the markdown output format: 'Markdown is cleaned and simplified, with headings, lists, links, and code blocks preserved. HTML tables are converted to markdown tables. Images are referenced by alt text + URL.'
• Add idempotency hint: 'This tool is idempotent, repeated calls with the same URL and options return the same markdown (results are cached). Safe to retry on transient failures.'

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

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry