Everyone has the same models. Your edge is context.

Oxagen graphs your business ontology and codebase into a typed knowledge graph. Agents get richer context, run on cheaper models, and ship evals that prove the delta.

MCP-native · Neo4j-backed · Workspace-scoped · Evaluations included

95%

Inference cost reduction

<50ms

p95 3-hop graph retrieval

2 graphs

Business ontology + codebase

Evals

Built in, not bolted on

The context layer is the margin

Models are commoditized. Context is not.

Every team on earth has access to GPT-4o, Claude, and Gemini. The teams that win aren't paying more per token — they're giving their agents richer context so they can pay less. A stateless model guesses what Order means in your system and which service owns it. An agent backed by your Oxagen graph knows — and you swap it out for Haiku, cut 95% of inference cost, and get a more accurate answer.

Business ontology

Entities, relationships, and domain schema auto-discovered from your data sources. Agents traverse your business model, not a flat vector index.

Code graph

Classes, functions, data models, and call graphs. Agents understand your implementation — not just your docs — so they stop hallucinating APIs that don't exist.

Built-in evals

Compare accuracy and cost per query across model tiers. See the delta. The graph is the argument; the evals are the proof your CTO needs to approve the infrastructure.

How it works

Graph. Query. Downgrade your model.



1

Step 1 · Graph

Graph your context

Connect your data sources and point at your codebase. Oxagen builds a typed ontology — business entities, relationships, code structure, all linked into one traversable graph.

Business entitiesCode graphRelationshipsEmbeddings
2

Step 2 · Query

Agents query the graph

Over MCP, your agents traverse the ontology instead of hallucinating. They know your domain schema, your service boundaries, your implementation details. No stateless prompt engineering.

MCP-nativeMulti-hopTyped SDKSSE stream
3

Step 3 · Downgrade

Run cheaper. Ship evals.

Context does the work the model was compensating for. Swap GPT-4o for Haiku, cut 95% of inference cost, measure the accuracy delta. The graph is the argument — the evals are the proof.

Model swapCost trackingAccuracy evalsPerformance delta

Context layer

The graph your agents should have been reading all along.

Your codebase has structure. Your business has a schema. Your agents should know both — not rediscover them in every prompt. Oxagen maps both into one typed, traversable graph so agents stop guessing and start knowing.

  • Business ontology: entities, relationships, and domain schema auto-discovered from connected sources
  • Code graph: classes, functions, data models, and call paths — agents understand your implementation, not just your docs
  • Hybrid semantic + graph traversal — sub-50ms multi-hop queries over context no RAG pipeline can assemble

Connections

Connect and forget. The ontology ingests continuously.

OAuth, APIs, MCP servers — wire a source in once and the graph stays current. Every agent you build downstream inherits a live, typed view of the underlying data without babysitting a pipeline.

  • Universal connectors for email, calendar, finance, storage, and code repositories
  • Agent tools and vector stores can push into the ontology over MCP
  • One typed pipeline per source class — no per-integration glue code

Agents

Shared memory for agent teams.

Stop giving each agent a stateless prompt. Agents share one ontology — same entities, same history, same reasoning. Deterministic entity IDs mean every agent refers to the same node across runs and sessions.

  • Claude, ChatGPT, and your own stack all plug in over the ontology MCP
  • Deterministic entity IDs — every agent refers to the same node across runs and sessions
  • Typed SDK + SSE streaming for reasoning traces, plus graph queries for precision lookups

Evaluations

See the delta. Justify the infrastructure.

Oxagen ships evaluations as a first-class feature — not an afterthought. Compare agent performance before and after graph context. Measure output quality, retrieval accuracy, and cost per query.

  • Side-by-side model comparison: accuracy with graph context vs. expensive stateless models
  • Cost per query tracked across model tiers — show your CTO the 95% inference cost reduction
  • Eval runs on every graph update — accuracy improves as the ontology sharpens

Security

Tenant-isolated. Agent-safe.

PostgreSQL row-level security, per-tenant encryption, and no training on your data. Ship agent workflows without leaking context across tenants — isolation primitives are wired into the graph, not bolted on later.

  • Row-level security enforces tenant isolation for every agent call — no implicit joins across tenants
  • OAuth tokens and graph payloads encrypted AES-256-GCM at rest; TLS 1.3 in transit
  • Your data is never used to train models; SOC 2, GDPR, HIPAA compliance roadmap in flight

From the blog

Engineering notes on agent infrastructure.

Read all posts

Run faster models. Get better results.

Graph your context. Let agents traverse it. Ship evals that prove the delta — in accuracy, in cost, in agent performance.

Get started freeRead the docs

No sales call required. Self-serve from install to first query.