gDown — The replicated graph database, native and AI-ready

gDown is the only component distributed independently for now. It is a replicated graph database speaking openCypher over HTTP and Bolt, compiled as a native image. It is AI-ready, built for AI assistants as well: an MCP server, vector and full-text search, and writes that cannot break anything.

Standards coverage

100.00 % of the openCypher TCK scenarios pass in the latest nightly run: 3,863 of 3,863.

Capabilities

What gDown does today, in the order an application meets it. Each statement comes from the project's documentation and tests, and limits are stated where they apply.

Graph

openCypher over Bolt 5.0 to 5.4, 5.6 to 5.8, 6.0 and 6.1 (the official Neo4j drivers, tested with Java and Python), HTTP and WebSocket; gDown's own test suite plays all 3,863 openCypher TCK scenarios, none skipped. Quantified path patterns and the ANY, ALL SHORTEST and SHORTEST k selectors, several databases, roles with privileges down to the property, and Raft replication (one group for every database by default, a group of its own for a database created with TOPOLOGY). Not yet: sharding, remote aliases, LDAP and SAML.

Vector search

CREATE VECTOR INDEX on a property that holds a list of numbers or a VECTOR value, then db.index.vector.queryNodes returns the k nearest neighbours by cosine or Euclidean similarity. The search is exact up to 2,000 vectors, then an approximate HNSW graph that follows updates. An optional filter (labels, property values, IN) is applied before the k are counted, and the rows returned feed a MATCH in the same statement. The index is held in memory and filled on first use.

Full-text search

CREATE FULLTEXT INDEX on one or several labels or relationship types and properties, with the analyzers standard, simple, whitespace, english (light stemming) and keyword. db.index.fulltext.queryNodes accepts terms, +required, -excluded, AND / OR / NOT, "exact phrase", prefix*, fuzzy~, property:term and parentheses, and scores by tf-idf. The index is held in memory and rebuilt when the database opens.

MCP server

POST /mcp speaks the Model Context Protocol. Its tools include get_schema, run_read_query, explain_query, profile_query, run_write_query with a dry run, and your saved queries. An assistant authenticates with a revocable API token or an account, so it has that account's roles and privileges, and a tool it may not use is not listed; in staged mode its writes become proposals that a reviewer promotes. For local clients such as Claude Code, gdown mcp-stdio bridges stdio to the endpoint. Not yet: OAuth specific to MCP, and server-to-client streams.

AI memory (gDown AIM)

gDown AIM (AI Memory) makes gDown the memory of an assistant. A client pushes the Markdown files and commits of a repository; gDown extracts typed facts (rules, pitfalls, decisions, plan items, routes, properties), links them to concepts and answers through memory_context, memory_ask and memory_about within a token budget, each answer citing its source. memory_note is the only write an assistant makes. The search combines concepts, full text and, when an embedding provider (Ollama or an OpenAI-compatible API) is configured, vectors. One repository per ingestion request.

Sleeping databases

UNLOAD DATABASE (or POST /db/<database>/unload) gives a database's memory back while it stays online: every member takes a checkpoint and closes its engine, and the first request reloads it from its files, with no change on the client side. With gluonify.gdown.database.idle-unload-seconds (off by default, 900 advised on the platform), a database that nobody has used for that long is unloaded by itself. Measured on a Mac (arm64, JDK 25): 99 MB of files, 110 MB given back, UNLOAD in 77 ms, first request in 82 ms; 982 MB of files, 191 MB given back, UNLOAD in 109 ms, first request in 549 ms. For comparison, Neo4j's STOP DATABASE (Enterprise) shuts a database down completely: it is offline until START DATABASE.

REST API generation

gDown can publish named Cypher queries as REST APIs. For each database, a query (a favorite) is analysed: gDown proposes the HTTP method, deduces the parameter and result types, and generates an OpenAPI 3.1 contract served under /api/<database>/<contract>/, with a Swagger UI to try it.

  • Security: the API is a resource server. It checks the bearer token (JWT) from the OpenID Connect provider you declare (issuer, audience, signature, expiry), then the roles and scopes each operation requires.
  • Least privilege: each operation runs with a database role chosen when the contract is written, not with the caller's own rights.
  • Limits: body size (413), request rate (429 with Retry-After), maximum rows, and NDJSON streaming for large results.
  • Writes: per operation, choose between accepted (202, no body), done (counts of created, updated and deleted) or query (the results of another favorite).
  • Replicated: providers, favorites and contracts live in the database, so they are replicated, backed up and dropped with it.
  • On Gluonify: Higgs, the orchestrator, can expose these contracts on your own domain, behind a Charm token and rate limits; nothing else of gDown is reachable from outside.

Go further

  • Quick start: Start gDown with Docker, run a first Cypher query over HTTP, then connect from Java or Python with the Neo4j driver and an API token.
  • Comparison: Compare gDown with Neo4j, Memgraph, FalkorDB and other graph databases, feature by feature, according to each product's documentation.
  • Performance: Same graph, same machine: gDown measured against Neo4j, Memgraph and others on latency, throughput, memory and LDBC SNB.

Community

Questions, ideas or a problem to share: the gDown channel is on the Gluonify.io Discord. Security vulnerabilities are never reported there, but privately to beacon@gluonify.io. Join #gdown on Discord