The orientation path is the point — list the graphs, read the labels and relationship types, sample the properties that actually appear on them, and only then write Cypher. Property keys ranked by frequency also expose naming drift, which is the usual reason a hand-written query silently returns nothing.
FalkorDB's own server for its graph database: list the graphs, learn a schemaless graph's shape by sampling it, then run OpenCypher against it.
- OpenCypher execution against any graph on the instance, with a read-only variant that always runs through GRAPH.RO_QUERY
- Parameterised queries: values travel in a params object and are referenced as $name, instead of being concatenated into the Cypher text
- Graph schema discovery — node labels, relationship types, and optionally the connection topology as source, relationship and target triples
- Property-level discovery: sample nodes of a label or relationships of a type and get their property keys ranked by frequency, with the sampled count reported alongside what you asked for
- Listing the graphs on the instance, creating nodes and relationships, and deleting a graph you no longer want
- Read-only enforcement where it matters: all three schema tools run read-only, so they are safe against a replica
A reachable FalkorDB instance and Node.js 18 or newer. Runs from npm as @falkordb/mcpserver, with FALKORDB_HOST and FALKORDB_PORT pointing at the database and FALKORDB_USERNAME and FALKORDB_PASSWORD when it is protected. Defaults to stdio; set MCP_TRANSPORT=http and MCP_PORT for a networked client, where MCP_API_KEY turns on API-key authentication. FALKORDB_DEFAULT_READONLY=true makes the whole server read-only, which is the setting for a replica.
One command — FALKORDB_HOST=localhost FALKORDB_PORT=6379 npx -y @falkordb/mcpserver