An ontology primer.
Plain-English primer. What an is, how the Oxagen graph works, and how read it over . Start here if you're new — jump to the reference docs when you're ready to build.
11+
Node Types
MCP
Native protocol
Neo4j
Canonical graph
AES-256
At rest + in transit
Why an ontology?
AI without context is just guessing
Most companies run on 15–25 systems — CRM, finance, docs, code, support tickets. Each one only knows its own slice. The Oxagen links them into one map your can walk.
An ontology isn't a database. It's a labeled map of the things in your workspace, the properties they have, and the links between them. Palantir built a $250B company on this idea for defense and intelligence data. Oxagen brings the same idea to the agent stack.
Cross-source reasoning
Ask questions that span the systems your business actually runs on. "Which enterprise contracts renew this quarter, who owns them, and which service handles their billing?" — one query, three sources, one answer.
Agents with context
Agents that don't just react — they understand. A new GitHub PR isn't just a diff; it's a change to a function the billing service depends on, touching three customers' workflows.
Cheaper models, same answers
Better context lets you run smaller, cheaper models. Swap Opus for Haiku. Cut 95% of inference cost. Prove the accuracy held on the same eval dashboard.
Plain-language graph search
Ask in English, get a typed answer. The graph knows which node a question is about — your agent doesn't have to guess which table or which prompt to write.
Built-in evals
Every graph update re-runs your eval suite. See accuracy go up and cost go down on the same dashboard. The graph is the argument; the evals are the proof.
Walled off by default
Your graph belongs to you. AES-256 encryption at rest, TLS 1.3 in transit, row-level security between customers. Your data is never used to train anyone's model.
See it in action
A stateless model guesses what Order means. An Oxagen-backed agent knows.
GPT-4o sees the word Order in your prompt and improvises. An agent reading your Oxagen graph walks straight to the typed , sees which service owns it, and finds the function that processes it. Now you can run the agent on — about 27× cheaper than — and the eval suite proves accuracy held.
How Oxagen Works
Connect. Ontology. Agents. Three steps to ship.
Wire in typed data sources once. The ontology builds a workspace-scoped Neo4j graph. Your agents query it over MCP, REST, or the typed SDK.
Connect
Plug in your data and your code.
Wire in the systems your business already runs on — GitHub, CRM, finance, docs, messaging. OAuth and signed tokens mean Oxagen never sees a raw password. One-time setup, continuous ingest.
Learn moreGraph
Two graphs, one shared map.
Oxagen builds a business ontology (customers, contracts, accounts) and a code graph (classes, functions, call paths). Both live in one workspace-scoped Neo4j graph your agents traverse.
Learn moreAgents
Read it over MCP. Run it on Haiku.
Your agents (Cursor, Claude Desktop, your own stack) read the graph over MCP. Better context means smaller, cheaper models — and built-in evals prove the accuracy held.
Learn moreTerminology
The key concepts behind Oxagen
Tap any term to learn more. These are the building blocks of the Oxagen ontology surface.
What ships today
The core platform
Everything below is live in production today.
Available now
Core platform features
- Connect GitHub, your CRM, your docs, and your finance system
- Auto-built business ontology + code graph
- MCP server so Cursor, Claude Desktop, and ChatGPT can read the graph
- Eval harness — score the same agent with and without the graph
- Typed SDK and OpenAPI spec for your own agents
- Row-level tenant isolation, AES-256 encryption, audit log
- Shareable artifacts: eval reports, account summaries, code diffs
- 200 free credits per month on the Free tier
For Developers
Build on the ontology
Jump directly into the API docs. Every card links straight to the developer reference.
You've read the primer. Now build on it.
Plug in your data. Plug in your code. Point your agents at the graph. Ship evals that prove cheaper models still get the right answer.
Self-serve from install to first query.