Neo4j

Executes Cypher queries against a Neo4j graph database with built-in schema introspection and validation. The MCP server parses incoming Cypher commands, validates them against the current database schema (nodes, relationships, properties), and returns structured results with type information. This prevents malformed queries from reaching the database and enables intelligent query suggestions based on available schema.

Automatically discovers and exposes the Neo4j database schema (node labels, relationship types, property keys, constraints, indexes) through MCP tools. The server queries Neo4j's internal schema metadata tables and constructs a machine-readable schema representation that LLMs can use to understand the graph structure without manual documentation. This enables context-aware query generation and prevents queries that reference non-existent entities.

Stores conversation history, entity relationships, and reasoning state in Neo4j as a persistent knowledge graph. Each conversation turn creates nodes for entities, intents, and decisions; relationships capture dependencies and temporal ordering. The MCP server retrieves relevant subgraphs on each turn to provide LLMs with long-term context without token overhead. This enables multi-turn conversations where the agent remembers prior decisions and can reason over accumulated knowledge.

Tracks intermediate reasoning steps, hypotheses, and decision branches as nodes and relationships in the graph. Each reasoning step creates a node with the step content, confidence, and dependencies; relationships capture causal chains and alternative paths. The MCP server enables agents to query the reasoning graph to backtrack, explore alternatives, or explain decisions. This provides transparency into agent reasoning and enables recovery from incorrect intermediate conclusions.

Accepts unstructured text (documents, conversation turns, web content) and constructs graph nodes and relationships by extracting entities and their relationships. The MCP server provides tools for entity extraction, relationship inference, and graph construction that LLMs can call iteratively. Extracted entities are deduplicated against existing graph nodes using fuzzy matching or semantic similarity, and new relationships are merged with existing ones to build a unified knowledge graph.

Integrates vector embeddings with graph structure to enable semantic search over entities and relationships. The MCP server stores embeddings for entity descriptions or properties, and provides a search tool that finds semantically similar entities using vector similarity while respecting graph structure (e.g., 'find entities similar to X that are connected to Y'). This combines the expressiveness of graph queries with the flexibility of semantic search.

Executes write operations (CREATE, UPDATE, DELETE) as ACID transactions, ensuring consistency even when multiple agents or processes modify the graph concurrently. The MCP server wraps Cypher write queries in Neo4j transactions, handles conflicts and retries, and provides rollback capabilities. This prevents data corruption and race conditions in multi-agent systems where multiple LLM instances may modify the same graph.

Executes complex graph traversal queries that match relationship patterns and navigate multi-hop paths. The MCP server provides tools for common traversal patterns (shortest path, all paths, k-hop neighbors) and allows custom Cypher patterns for domain-specific queries. This enables agents to discover indirect relationships, identify clusters, and reason over graph topology.

Defines and enforces constraints on graph data (uniqueness, existence, type constraints) to maintain data quality. The MCP server exposes Neo4j's constraint system, allowing agents to define constraints declaratively and validate data before writes. Constraints prevent invalid states (e.g., duplicate entities, missing required properties) and provide clear error messages when violations occur.