Server has 6 tools with explicit schemas and descriptions. Naming follows verb_noun convention (search_, capture_, list_, update_, run_, get_). Descriptions are present but vary in quality (72-180 chars). All tools have typed input schemas with parameter descriptions. However, output schemas are not documented in the source, parameter constraints lack depth, and error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Composition is reasonable but some tools combine multiple concerns (capture_conversation does classification + storage).

• Document output schemas for all 6 tools. Example: 'Returns array of {id: string, title: string, url: string, score?: number}' for search_knowledge. Include pagination info (total_count, next_cursor) for list_records.
• Add tool annotations: mark capture_conversation, update_record_status, run_follow_ups with destructiveHint=true; mark search_knowledge, list_records, get_widget with readOnlyHint=true. Use @modelcontextprotocol/sdk registerAppTool annotations.
• Split capture_conversation into two tools: capture_conversation (stores message) and optionally classify_conversation (AI classification). Allows agent to capture without classification if desired.
• Enhance parameter descriptions with explicit constraints. E.g., 'limit: Maximum results (1 - 100, default 20)' and 'provider: Classification engine (auto|heuristic|openai|anthropic; requires OPENAI_API_KEY if openai selected)'.
• Add recovery guidance to descriptions. E.g., search_knowledge: 'If no results, try shorter query or check ALLOWED_DOC_HOSTS config.' capture_conversation: 'If classification fails, verify provider credentials and try provider=heuristic.'