When to use it#

The MCP servers are for letting an LLM agent use AgensGraph directly through tools — for example in Claude Desktop — with no application code of your own. They call no LLM themselves, so they are deterministic and easy to test.

Good fits

  • Chat with a graph. Let Claude Desktop (or any MCP client) answer natural-language questions by calling get_agensgraph_schema and the paginated, read-only read_agensgraph_cypher against your graph.

  • Conversational memory. Give an assistant a knowledge-graph memory it grows with create_entities / add_observations and recalls with search_memories / read_graph across a conversation.

  • Schema-design assistant. Use the create_new_data_model prompt to design a graph data model, render a Mermaid diagram, and generate constraint + ingest Cypher — then load it through the Cypher server’s write tool (started with --allow-graph-ddl for the constraint statements).

  • Query tuning. Ask why a query is slow: explain_agensgraph_cypher shows the plan, recommend_property_indexes suggests indexes and rewrites, and agensgraph_health reports cache hit ratio, unused indexes and vacuum backlog.

  • Safe, read-only exploration. Expose a database for querying with a server-wide read-only mode and paginated results, so an agent can neither flood the context nor modify data. Connect as an ordinary role, not a superuser or a member of pg_execute_server_program; the Cypher and memory servers refuse to start as one unless --allow-server-programs is given.

  • Multiple graphs and transports. Run a namespaced server per graph, over stdio or Streamable HTTP, side by side.

How it compares

  • Pick the MCP servers when you are enabling an agent or assistant to work with AgensGraph, rather than building a Python application.

  • Pick LangChain, LlamaIndex, LightRAG, or cognee for in-application graph and vector storage in your own code.

Consider an alternative when

  • You are building a Python application — use one of the library integrations for full control in code.