[Guide]
Best open source graph databases in 2026
By the Omnigraph team · Last checked
Twelve graph databases ship their engine under an OSI-approved license and are actively maintained. Neo4j Community Edition is still the default for Cypher and a large ecosystem. Dgraph, NebulaGraph and JanusGraph scale out across machines, LadybugDB embeds in your process, and Omnigraph and TerminusDB are the two that let you branch, merge and read the graph as it was.
Six more engines that often appear on “open source” lists (Memgraph, FalkorDB, ArangoDB, SurrealDB, Fluree, HydraDB) are source-available, not open source: their licenses restrict offering the database as a service to others.
We build Omnigraph, one of the engines on this list. Every claim here links to a public source, the comparison is a script you can rerun, and Omnigraph’s weak spots are listed with its strengths.
[Picks]
Best for
Branches, merges and reads at an earlier commit
Omnigraph, TerminusDB
Cypher, the largest ecosystem, the most tutorials
Neo4j Community Edition
An embedded graph in your process
LadybugDB
Scale-out across many machines
Dgraph, NebulaGraph, JanusGraph
Strict typing and rich modeling
TypeDB
A graph inside the PostgreSQL you run
Apache AGE
Graph plus documents in one engine
ArcadeDB, OrientDB
A graph on S3-compatible object storage
Omnigraph, HelixDB
[Method]
How we chose and checked
Inclusion rules
- The engine itself is under an OSI-approved license, read from the repo’s LICENSE file. A permissive SDK on top of a source-available server doesn’t count.
- It’s a database you run, not a query layer over another store or a framework (so PuppyGraph and Apache TinkerPop are out).
- It’s active: a release or commits in the last 12 months. That drops Cayley (Apache-2.0, last release 2019). Kùzu was archived in October 2025; its maintained fork, LadybugDB, is on the list.
OSI-approved means on the Open Source Initiative’s license list.
The runnable test
We loaded the same small dataset into all twelve engines and ran the same four steps, each in a fresh Docker container (LadybugDB as a pinned Python package), with the version pinned and logged.
- Create the schema each engine’s docs recommend for typed data, then load 12 people, 3 companies, 5 projects and the edges between them.
- Run a two-hop query: which projects do Acme’s employees contribute to? The script checks the answer against the dataset (
Atlas, Beacon, Comet). - Try two bad writes through the normal write path: a person with
age: "forty", then one with an undeclarednickname. Record whether each is rejected, by default and again in the engine’s strict mode where it documents one. - Where the engine documents branches or versioned reads: branch, write on the branch, check main is untouched, merge, then read the commit from before the merge.
Scripts, dataset and full logs: ModernRelay/graph-database-comparison. Rerun it with python run_all.py.
Sources
Every cell in the comparison table links to a public source (docs, release notes or code) and carries the date it was last checked.
Keeping it current
A script checks every repo’s license file, archived flag, latest release and activity on the first Monday of each month. When something moves, the table changes and the changelog says what changed.
[Run]
What the run showed
Every engine answered the two-hop query correctly. The differences show up when a write breaks the schema and when you want a branch.
| Engine (version run) | Two-hop query | Rejects age: "forty" | Rejects undeclared nickname | Strict mode | Branch, merge, older commit |
|---|---|---|---|---|---|
| Omnigraph v0.15.0 | pass | rejected | rejected | not needed | pass, pass, pass |
| TerminusDB v12.0.9 | pass | rejected | rejected | not needed | pass, pass, pass |
| Neo4j CE 2026.10.2 | pass | accepted | accepted | refused in Community Edition | not available |
| Dgraph v25.5.1 | pass | rejected | accepted | rejects both (--limit mutations=strict) | not available |
| NebulaGraph 3.8.0 | pass | rejected | rejected | not needed | not available |
| HelixDB server v0.0.15 | pass | accepted | accepted | none documented | not available |
| JanusGraph 1.2.0 | pass | rejected | accepted | rejects both (schema.default=none, schema.constraints=true) | not available |
| Apache AGE 1.8.0 (PG 18) | pass | accepted | accepted | none documented | not available |
| TypeDB CE 3.14.2 | pass | rejected | rejected | not needed | not available |
| LadybugDB 0.22.3 | pass | rejected | rejected | not needed | not available |
| ArcadeDB 26.11.1 | pass | rejected | accepted | none documented | not available |
| OrientDB 3.2.58 | pass | rejected | accepted | rejects both (STRICTMODE TRUE on the class) | not available |
Every command and its output is in each engine’s run log in ModernRelay/graph-database-comparison.
“Not available” means the engine documents no branches or versioned reads, so there was nothing to run; the comparison table shows what each offers instead (usually backups you restore, not query).
Accepting undeclared properties lets you change the data model without a migration, and moves validation into your application. The test shows which engines reject a bad write without extra setup.
[Compare]
Open source graph database comparison
Checked November 11, 2026. Stars from GitHub on the same day.
| Engine | License | Query languages | ACID | Branch and merge | Versioned reads | Object storage | Larger than RAM | Scale-out | Hybrid search | Embedded | MCP server | Stars |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Omnigraph | MIT | GQ (own) | 2.0k | |||||||||
| TerminusDB | Apache-2.0 | WOQL, GraphQL | 3.5k | |||||||||
| Neo4j CE | GPL-3.0 | Cypher, GQL | 17.4k | |||||||||
| Dgraph | Apache-2.0 | DQL, GraphQL | 21.8k | |||||||||
| NebulaGraph | Apache-2.0 | nGQL, openCypher | 12.5k | |||||||||
| HelixDB | Apache-2.0 | Helix query DSL (own), openCypher | 6.3k | |||||||||
| JanusGraph | Apache-2.0 | Gremlin | 5.9k | |||||||||
| Apache AGE | Apache-2.0 | openCypher, SQL | 4.9k | |||||||||
| TypeDB | MPL-2.0 | TypeQL | 4.5k | |||||||||
| LadybugDB | MIT | Cypher | 1.9k | |||||||||
| ArcadeDB | Apache-2.0 | SQL, openCypher, Gremlin, GraphQL, MongoDB QL | 1.3k | |||||||||
| OrientDB | Apache-2.0 | SQL, Gremlin | 5.0k |
✓ yes · ◐ partial · – no. “Hybrid search ✓” means graph, vector and full-text search in one query with built-in rank fusion; ◐ means some of those in one query, without fusion. Hover a mark for its note; each links to its source.
[Engines]
The engines
Omnigraph
MIT · Rust · 2.0k stars · Run log (v0.15.0)
- What it is
- A typed property-graph database in Rust, first released in April 2026. Schemas are declared in
.pgfiles and queries are named, typed.gqfiles. The graph is stored as Lance datasets on local disk or S3-compatible storage, GCS or R2 (Azure in preview). - Best for
- Agent and team workflows where changes should be reviewable: a branch per agent or task, three-way merge across the whole graph, and queries against any earlier commit with
--snapshot. Graph, vector and full-text search combine in one query with Reciprocal Rank Fusion, andomnigraph-serverhas an MCP endpoint built in. - Watch out for
- No horizontal scaling yet: one server holds the writer lease, and clients write through it on isolated branches. No LangChain or LlamaIndex integration (MCP, a TypeScript SDK and HTTP only). It’s the youngest project here and has the fewest stars. Omnigraph Cloud is in private preview.
+ All 24 sourced facts
TerminusDB
Apache-2.0 · Prolog, Rust · 3.5k stars · Run log (v12.0.9)
- What it is
- A document-graph database that stores typed JSON documents as an RDF graph under a closed-world schema. The engine core is SWI-Prolog, with the storage layer in Rust. Every write is an immutable commit.
- Best for
- Versioned data with a schema: branches, field-level three-way merge, diff, and queries at any commit. It has the broadest git-style feature set on this list (rebase, squash, push, pull and clone as well), and a migration API that renames fields and casts types.
- Watch out for
- The docs call it an in-memory database: data persists on disk, but queries run in memory, so the working set has to fit. Single node (clustering is in the commercial Enterprise edition). Vector, full-text and hybrid search and an MCP server arrive in the 12.1 release candidate, not the current stable release.
+ All 24 sourced facts
Neo4j Community Edition
GPL-3.0 · Java, Scala · 17.4k stars · Run log (2026.10.2)
- What it is
- The property-graph database that introduced Cypher, written in Java and Scala. Version 2026.10 uses calendar versioning with Cypher 25, which Neo4j says supports most mandatory GQL features.
- Best for
- Cypher with the largest set of drivers, tools and tutorials. Vector and Lucene full-text indexes are in Community Edition, there’s an official MCP server, and LangChain, LlamaIndex and others integrate with it.
- Watch out for
- Community Edition is single-instance: clustering, online backup, and existence, type and key constraints are Enterprise-only. In the run, CE accepted both bad writes and refused to create a property type constraint.
+ All 24 sourced facts
Dgraph
Apache-2.0 · Go · 21.8k stars · Run log (v25.5.1)
- What it is
- A distributed graph database in Go with DQL and a native GraphQL layer, sharded by predicate over the Badger key-value store. Since v25, formerly enterprise features (CDC, ACLs, binary backups) ship in the open-source build.
- Best for
- Scale-out with distributed ACID transactions (snapshot isolation, Jepsen-tested) and a GraphQL API without a separate server. Vector search with HNSW and an MCP endpoint are built in.
- Watch out for
- By default, unknown predicates are created on write: the run accepted the undeclared
nicknameuntil Dgraph ran in strict mode. No managed cloud (the current owner, Istari Digital, says none is planned) and no embedded mode.
+ All 24 sourced facts
NebulaGraph
Apache-2.0 · C++ · 12.5k stars · Run log (3.8.0)
- What it is
- A distributed C++ property-graph database with separate graph, meta and storage services over RocksDB and Raft, queried with nGQL and an openCypher subset.
- Best for
- Large graphs hash-sharded across storage nodes, with a strongly typed schema enforced by default. Both bad writes were rejected.
- Watch out for
- The docs state that nGQL doesn’t support transactions. The latest open-source release is 3.8.0 from May 2024; new features (GQL, vector search) go to the closed Enterprise 5.x line. Full-text search needs an external Elasticsearch cluster, and it’s three services to run, not one.
+ All 24 sourced facts
HelixDB
Apache-2.0 · Rust · 6.3k stars · Run log (server v0.0.15)
- What it is
- A Rust graph database (Apache-2.0 since May 2026, AGPL-3.0 before) with vector and BM25 search on SlateDB. It runs on S3-compatible storage, local disk or in memory, as a server or embedded in Rust, TypeScript or Python.
- Best for
- A graph plus vector and keyword search on object storage, with traversal, ANN and BM25 in one transaction. Besides its own query DSL, the server answers openCypher at
/v2/cypher; in the run it returned the same three projects. - Watch out for
- Version 3 dropped declared schemas, so any label or property is accepted, and the run accepted both bad writes. No built-in rank fusion. The official MCP endpoint and auto-scaling readers are Helix Cloud only; the self-hosted server is a single node.
+ All 24 sourced facts
JanusGraph
Apache-2.0 · Java · 5.9k stars · Run log (1.2.0)
- What it is
- A distributed Java property-graph database under the Linux Foundation, queried with Gremlin. It’s a layer over a storage backend you choose (Cassandra, ScyllaDB, HBase, Bigtable, BerkeleyDB JE or in-memory), plus an optional index backend for full-text search.
- Best for
- Teams that already run Cassandra, ScyllaDB or HBase and want a graph on top that scales out with them. No company, paid edition or license key.
- Watch out for
- The docs say transactions aren’t necessarily ACID: they can be on BerkeleyDB, and generally aren’t on Cassandra or HBase. No vector search in 1.2.0. By default it creates property keys on first use, so the run accepted
nicknameuntil the strict settings were on. The run used BerkeleyDB JE with Lucene.
+ All 24 sourced facts
Apache AGE
- What it is
- A PostgreSQL extension that adds openCypher over a property graph stored in ordinary PostgreSQL tables. Cypher runs inside SQL through the
cypher()function. - Best for
- Adding graph queries to a PostgreSQL you already operate. Transactions, MVCC, backups and replication are PostgreSQL’s, and SQL and Cypher mix in one statement. Vector search comes from pgvector, whose functions work inside Cypher.
- Watch out for
- Properties are a free JSON-like map with no type checks, so both bad writes went in. No sharding: the AGE FAQ says its tables can’t be distributed with Citus yet. Full-text search is SQL-side, not in Cypher.
+ All 24 sourced facts
TypeDB
MPL-2.0 · Rust · 4.5k stars · Run log (CE 3.14.2)
- What it is
- A strongly typed database with its own type-theoretic model (entities, n-ary relations and attributes with inheritance) and the TypeQL language. Rewritten in Rust for 3.0, on RocksDB.
- Best for
- Domains where the model is the hard part: relations between more than two things, type hierarchies, and every write validated against the schema.
- Watch out for
- Community Edition is single-server; replication is Cloud and Enterprise only. No full-text index, and native vector search is an alpha for invited testers. No LangChain or LlamaIndex package.
+ All 24 sourced facts
LadybugDB
MIT · C++ · 1.9k stars · Run log (0.22.3)
- What it is
- The community fork of Kùzu, started in October 2025 after Kùzu’s repo was archived. An embedded, columnar, single-file graph database with Cypher; it also runs in Wasm.
- Best for
- Analytics and local graphs inside your application, with no server to run. Graphs are strictly typed by default, and both bad writes were rejected.
- Watch out for
- One write transaction at a time, and one read-write process per database file. No scale-out. Vector and full-text search are official extensions you
INSTALLandLOAD.
+ All 24 sourced facts
ArcadeDB
Apache-2.0 · Java · 1.3k stars · Run log (26.11.1)
- What it is
- A multi-model database in Java (graph, document, key/value, time series, vector) from the founder of OrientDB. One edition, with no feature gating. It speaks SQL, openCypher, Gremlin, GraphQL and the MongoDB and Redis protocols, plus Bolt and the Postgres wire protocol.
- Best for
- One engine for several models and query languages, with ACID transactions, Raft replication in the open-source edition, vector and full-text indexes and a built-in MCP server.
- Watch out for
- Declared properties are validated, but undeclared ones are accepted, and we found no stricter mode in the docs: the run rejected
age: "forty"and acceptednickname. Replicas hold full copies, with no sharding.
+ All 24 sourced facts
OrientDB
Apache-2.0 · Java · 5.0k stars · Run log (3.2.58)
- What it is
- A Java multi-model database (graph, document, key-value, object) queried with an extended SQL with MATCH patterns, and Gremlin through a TinkerPop package. After SAP ended its maintenance in 2023, it continues as a community project with roughly monthly 3.2.x releases.
- Best for
- Existing OrientDB users, and embedded or server multi-model use in Java with Lucene full-text search. The former Enterprise Agent (incremental backup, cluster management) is now free under Apache-2.0.
- Watch out for
- Classes are schema-less or hybrid by default; undeclared properties are rejected only in strict mode. No vector index. Sharding is manual and the docs flag its limitations. Version 4.0 is in development with no beta yet.
+ All 24 sourced facts
[Licenses]
Often called open source: six source-available engines
You can read their code, and most can be self-hosted for free, but their licenses aren’t OSI-approved and restrict some uses, usually offering the database as a service.
| Engine | License | What it means in practice |
|---|---|---|
| Memgraph | BSL-1.1 | Converts to Apache-2.0 after four years. Community needs no key; RBAC, multi-tenancy and automatic failover need an Enterprise license. |
| FalkorDB | SSPL-1.0 | Self-host the Redis module with no key; commercial license on request. |
| ArangoDB | BSL-1.1 | Source-available since 3.12: internal production allowed, no competing commercial offering. The free prebuilt binaries use a separate Community License: non-commercial, read-only past 100 GiB. |
| SurrealDB | BSL-1.1 | Converts to Apache-2.0 on 1 January 2030; no third-party database-as-a-service. SDKs are Apache-2.0. |
| Fluree | BSL-1.1 | Free in production except as a hosted database service; each release converts to Apache-2.0 after four years. Has branching, merge and time travel. |
| HydraDB | Elastic-2.0 | AGPL-3.0 until 3 November 2026. The license bars offering it as a managed service; vector search, BM25 and the memory API are Cloud only. |
Watched, not listed: Kùzu (Archived in Oct 2025; LadybugDB is the maintained fork. Mentioned on the page, not listed.) Cayley (Apache-2.0 but last release 2019. Fails the activity rule; listed again if it ships.) Not covered: managed and proprietary engines such as Amazon Neptune, TigerGraph and Google Spanner Graph.
[FAQ]
Questions
What is an open source graph database?
A graph database stores data as nodes and the relationships between them, and answers queries by following those relationships, in languages such as Cypher, Gremlin or GQL. It’s open source when its engine is released under a license the Open Source Initiative has approved, such as Apache-2.0, MIT, GPL-3.0 or MPL-2.0. Licenses like BSL, SSPL and the Elastic License let you read the code but restrict use, so they’re source-available. Check the LICENSE file of the engine itself, not just the SDKs: several vendors use permissive licenses for clients and a restrictive one for the server.
Is Neo4j open source?
Neo4j Community Edition is, under GPL-3.0. Enterprise Edition is commercial and closed source, and several features are Enterprise-only, including clustering, online backup and property type constraints. In our run, Community Edition refused to create a type constraint.
What’s the best open source graph database for AI agents?
Pick by what the agents do with the graph. If several agents write and their changes should be reviewed or rolled back, you want branches and versioned reads: Omnigraph and TerminusDB are the open-source engines with both. For the widest framework support, Neo4j Community Edition has LangChain and LlamaIndex integrations and an official MCP server. For a graph inside your own process, LadybugDB.
Can I run a graph database on S3?
Yes. Among open-source engines, Omnigraph stores the graph as Lance datasets on S3-compatible storage, GCS or R2, and HelixDB uses SlateDB on S3-compatible storage. LadybugDB can keep Parquet tables on S3, but the database itself is a local file. Fluree and HydraDB also run on object storage, under source-available licenses.
Which open source graph databases can run embedded?
Seven of the twelve run inside your process. LadybugDB is an in-process library with a single-file database, and also runs in Wasm. Omnigraph ships a Rust engine crate, and its CLI runs in-process against a local graph. HelixDB has embedded runtimes for Rust, TypeScript and Python. Neo4j Community Edition, JanusGraph, ArcadeDB and OrientDB embed in a JVM process (ArcadeDB also has Python bindings). Dgraph, NebulaGraph, TypeDB, TerminusDB and Apache AGE need a server.
What replaced Kùzu?
Kùzu’s repository was archived in October 2025. LadybugDB is the maintained community fork, under MIT, with the same embedded, columnar design and Cypher.
[Changelog]
What changed
First version. 12 open-source engines, 6 source-available. HydraDB listed as source-available after its 3 November relicense from AGPL-3.0 to Elastic License 2.0.