Server has 5 well-named tools with consistent verb_noun patterns (search_patents, get_biblio, get_family, get_abstract, get_claims). All tools have descriptions (avg 95 chars, within 10-1024 range). Input schemas are present with type declarations and required fields. However, parameter descriptions are minimal (e.g., 'Patent number, epodoc format' lacks format constraints, examples, or validation guidance). Output schemas are completely undocumented, LLMs cannot predict response structure or plan downstream calls. Error handling exists but lacks recovery guidance (e.g., 'EPO: not found' doesn't suggest alternatives). No tool annotations (readOnlyHint, idempotentHint). Credentials passed via _apiKey parameter violates secret-injection pattern, should use server-side environment injection.

• Document output schemas for all tools. Example for search_patents: 'Returns {total_results: number, results: [{publication_number: string, title: string}], range: string}'. This enables LLMs to extract and chain results.
• Move _apiKey to server-side environment variable (EPO_CONSUMER_KEY, EPO_CONSUMER_SECRET). Remove from tool parameters. Initialize token cache at server startup, not per-call.
• Add format constraints to parameter descriptions. E.g., 'Patent number in epodoc format (e.g., EP1234567, EP-1234567B1). Must be 2 letters + 7 digits + optional suffix.'
• Enhance error messages with recovery guidance. E.g., 'EPO: not found (EP1234567). Verify the patent number format or try search_patents() to find the correct number.'
• Add tool annotations: readOnlyHint=true and idempotentHint=true to all tools. This signals to agents that calls are safe to retry without side effects.