The BNFR MCP server has 4 tools with mixed quality. Naming is clear and action-oriented (all start with verbs: get_*, search_). Descriptions are present but vary in detail. Input schemas are visible in the source code with proper types and descriptions. However, output schemas are largely undocumented, and error handling is minimal. The server uses FastMCP with HTTP transport and proper tool registration via decorators. The two time tools (get_time_in_london, get_time_in_new_york) are trivial and lack meaningful descriptions. The search_knowledge_base tool has a reasonable description and proper parameter types, but the joke tool lacks parameter descriptions. Overall, the server exhibits basic competence but falls short of production-grade tool design, most tools lack output schema documentation, error handling is absent, and parameter descriptions are inconsistent.

• Add return type documentation to each tool. For example, search_knowledge_base should document: 'Returns a JSON array with fields: [{ content: string, metadata: { filename, repository, ... } }]'. For time tools, document: 'Returns a string formatted as YYYY-MM-DD HH:MM:SS ZZZ'.
• Expand descriptions of time tools to 50 - 150 characters with context. Example: 'Get the current time in London, UK. Use this to coordinate meetings with UK-based teams or check if it is business hours in London.'
• Add constraints to the search_knowledge_base 'filename' parameter schema: { 'type': 'string', 'pattern': '^[\w.-]+$', 'description': 'Filter by filename (e.g., README.md, config.yaml). Only alphanumeric, hyphens, dots, and underscores allowed.' }
• Implement error handling in the search_tools.py:search_knowledge_base() function. If results are empty, return: { 'error': 'No documents found', 'suggestion': 'Try a broader query or check that repositories are indexed in the knowledge base' }. This guides LLM retry behavior.
• Add per-tool error classification. For search_knowledge_base, document in the description: 'Returns results or an error message. If 'No documents found', the agent should suggest rephrasing the query or listing available repositories.'
• Document that get_random_joke takes no input parameters and returns an object with 'setup' (string) and 'punchline' (string) fields. Update description: 'Returns a random programming joke with setup and punchline. No parameters required. Use when the user asks for a joke or to inject humor into a conversation.'