Neo4j MCP server has 20 tools with consistent naming (verb_noun pattern) and basic schemas. However, descriptions are generic and lack LLM-optimization guidance. Most parameter descriptions are minimal (10-30 chars), falling below the 72-char baseline. Output schemas are undocumented, LLMs cannot infer what fields to expect from responses. Error handling is absent; tool implementations catch exceptions but return generic error text without recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification in the metadata. Security concerns: write/destructive tools lack confirmation patterns or dry-run support.

• Add tool annotations: set readOnlyHint=true for all READ_ONLY tools (get_*, read_*, find_*, explain_*), destructiveHint=true for drop_* tools. This enables LLMs to reason about safety.
• Document output schemas for all tools. E.g., get_neo4j_schema should declare it returns {nodes: [{label, properties: [{name, type}]}, relationships: [{type, properties}]}. Use structured TextContent or define a custom response type.
• Implement confirmation pattern for destructive operations. Before drop_index/drop_constraint, return a result with 'input_required' asking user to confirm. This prevents accidental deletions.
• Enhance error messages with recovery guidance. E.g., 'Invalid Cypher syntax: {error}. Check query for typos or use explain_neo4j_cypher to validate.' Categorize errors as retryable vs user-fixable.
• Expand parameter descriptions to 50-100 chars with format/constraint hints. E.g., 'The Cypher query to execute (must be a valid read-only query; use MATCH, RETURN, WHERE, LIMIT)' for read_neo4j_cypher.

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

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