Invana

An ontology-governed operating system for humans and AI agents.

Invana is open source. People and agents work on the same knowledge graph, and the ontology governs both of them.

Core features

Invana is in beta. ✅ available · 🟡 usable, with parts still being built.

Feature	What you get	Status
Self-hosted	One Docker Compose stack on your own machines. Your graph and your keys stay with you, and with Ollama the LLM can run on your network too	✅
Ontology-governed	Models give the data its meaning, and stitches join models from different sources	✅
Guardrails, worlds and rules	Guardrails bound what any run may read, use or send; worlds narrow them per question; rules block, warn or report	🟡
Roles and access	Graph roles are enforced on every route, and search shows only what you can open	✅
Grounded answers	Every answer traces from the LLM to the query, the record and the dataset. When the graph can't answer, Invana says so	✅
Decisions with a trail	Projects keep the evidence for and against, and decisions are append-only	✅
Simulation	Run a simulator on a project branch over pinned dataset versions, reproducible from its seed	✅
Datasets and provenance	Origin, versions, licence and sensitivity on every dataset; a retraction flags everything that cites it	🟡
Agents	Plans generated from each agent's instructions, checked by evals before they publish, and bounded by token budgets	🟡
Plugins	Every task comes from a plugin. Core plugins are always on; tool, MCP and simulator plugins extend what agents can do (below)	🟡
Runs, schedules and approvals	One journal of every run, schedules with delivery, and a person's approval before anything leaves	🟡
Change log	Every versioned thing has a history; compare any two versions and see the effect after	🟡

Deploy

Invana has no published release yet, so you deploy it from source. You need Docker with Docker Compose.

git clone https://github.com/invana/invana.git
cd invana
docker compose up -d

On every start, a one-off engine-migrate container applies pending migrations, then the engine starts. No user exists until you create one.

Create the first users

1. Create the first install Admin. The command asks for a username, a name, an optional email and a password; nothing is filled in for you:

   docker compose exec engine uv run invana users create --superuser

2. Open Studio at http://localhost:8300 and sign in as that Admin.

3. Add the people who will use it. Each command asks for the password:

   docker compose exec engine uv run invana users create --username user --email user@invana.io

   Add --superuser for another install Admin. Graph roles are given per Graph, in Studio.

4. Reset a password when someone needs it:

   docker compose exec engine uv run invana users update-password --user user

What is running

Studio	http://localhost:8300
Engine API	http://localhost:8200
Graph database	Neo4j starts by default. Memgraph, JanusGraph and ArcadeDB are Compose profiles, for example docker compose --profile memgraph up -d
Sample data	The airways demo: models, records, stitches and guardrails, step by step

This is the development stack. Before anyone else can reach it, set INVANA_SECRET_KEY and INVANA_ENCRYPTION_KEY (CONTRIBUTING.md › Environment variables) and use strong passwords for every account.

Integrations

Graph databases

The graph backend is switchable: each Graph connects to the database you choose, and models, governance and the products work the same on any of them. Invana speaks both Cypher and Gremlin, so its generic connectors can reach other Cypher or Gremlin databases too. The ones below have their own tested connector.

Database	Query language	Connector
Neo4j	Cypher	invana-neo4j
Memgraph	Cypher	invana-memgraph
ArcadeDB	Cypher, Gremlin	invana-arcadedb
JanusGraph	Gremlin	invana-janusgraph
TinkerGraph	Gremlin	invana-tinkergraph
Amazon Neptune	Gremlin	planned

LLM providers

Provider	How it connects
Anthropic (Claude)	The official Anthropic API
Claude Agent SDK	Claude through a local Claude Code install, with an API key or a Claude subscription token
OpenAI	The OpenAI API
OpenAI-compatible servers	Any /chat/completions endpoint, such as vLLM, LM Studio or LocalAI
Ollama	Local models, no API key

Data and standards

Format	Use
SKOS	Import a vocabulary as a standard
RDFS, OWL (subset), schema.org	Import an ontology as a draft model
JSON-LD, Turtle, RDF/XML	Read those ontology files
OpenTelemetry	Traces, logs and metrics, for example into HyperDX

Plugins

Everything a run does is a task from a plugin. An install Admin enables tool and MCP plugins; core plugins are always on.

Plugin	Type	What it does	Status
Graph read	core	Run queries, count types, resolve and expand records, preview stitches	✅
Graph write	core	Write and bulk-load records; apply, commit and withdraw stitches	✅
Ingest	core	Check a bundle, snapshot its model, validate records before they land	✅
LLM	core	Understand a question, plan the work, turn a thought into a query, draft plans and models	✅
Schema write	core	Propose, validate and migrate model changes	✅
Work write	core	Spawn agents, delegate to them, create tasks	✅
Pure	core	Validate queries, verify results, shape results for the canvas, with no side effects	✅
Web	tool	Search, fetch, and download-and-extract. Ships the collect a dataset skill	✅
Mail	tool	Read a mailbox over IMAP; send over SMTP only after a person approves	✅
Science sources	tool	PubMed search, UniProt and PubChem lookups	✅
MCP	mcp	One plugin per MCP server an Admin adds; each of its tools becomes a task	🟡
Genomic prediction	simulator	Predict trait outcomes for a cross, run on a project branch over pinned datasets	✅
Plugin SDK	sdk	Manifests, the task context, errors and the simulator contract, for writing a plugin. Plugins ship inside the engine today	🟡

The products

Product	For	Job
Datasets	researchers, curators	Gather the data: reusable, labelled by origin, checked against standards
Modeller	modellers, compliance leads	Give it meaning: models, stitches, worlds, guardrails, rules
Explorer	analysts	Ask and look: answers drawn on a board, each claim citing its record
Projects	anyone deciding something	Pursue a question: gather evidence for and against, decide, keep the trail
Agents	builders	Build, tune and watch the agents that do the legwork

Contributing

CONTRIBUTING.md covers the development setup, how a feature goes from the feature index to code, and how to raise a pull request.