VKLP MCP server has 5 tools with clear verb-noun naming (search_knowledge, get_claim, propose_claim, verify_ledger, service_info). All tools have descriptions (10 - 200 chars, within baseline). Input schemas are present and typed for all tools. However, parameter descriptions are minimal or missing context about constraints, formats, and error recovery. Output schemas are not documented, callers must infer structure from code. Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. The propose_claim tool accepts JSON-encoded evidence as a string rather than structured input, forcing LLM serialization and increasing error surface. No tool declares permissions or security implications. Overall: functional but below production baseline for agent-facing tools.

• Document output schemas for all tools. E.g., search_knowledge returns {hits: [{claim_id, status, consensus_score, ...}], total_count}. Include field types and meanings.
• Refactor propose_claim to accept evidence as a structured array parameter (type: array of objects with 'source', 'text', 'confidence' fields) instead of JSON string.
• Add parameter constraints: top_k (1 - 100, default 6), language (enum: en, de, fr, ...), domain (enum: general, science, law, ...). Document in descriptions.
• Add error handling: document when tools return empty results vs errors. E.g., 'get_claim returns {found: false, record: null} if claim_id does not exist. Call search_knowledge to discover valid claim IDs.'
• Add tool annotations via @mcp.tool(readOnlyHint=True) for search_knowledge, get_claim, verify_ledger, service_info. Mark propose_claim with destructiveHint=True.