How the business works is scattered
It lives across documents, systems and people. Turning it into a governed, ontology-grounded graph is slow, manual and stale on arrival.
01What is TrustGraph?
An operating graph and a policy runtime, connected.
TrustGraph is an open-source platform that turns enterprise knowledge, systems, and logic into an ontology-grounded operating graph, then uses that graph to govern AI agents at the moment they act. Every proposed action is checked against business meaning, authority and evidence, and is allowed, paused or denied before it executes.
Each decision is stored with the exact evidence and policy version behind it, as W3C PROV-O provenance you can query later.
02The problem
Agents have credentials. They don't have context or limits.
Nothing governs the moment of action
An agent with API credentials can execute regardless of contract, authority or evidence. Guardrails filter text; they don't decide whether a payment should go out.
Enterprises need both, connected: a living graph of how the business operates, and a runtime that enforces it on every agent action.
03How it works
The semantic layer sits inside the execution loop
Agents can reason probabilistically. Their actions are decided deterministically.
Business knowledge
Ontologies and facts in a graph
Agent proposes
An LLM plans an action
Semantic policy check
Meaning, authority, evidence
Decision
ALLOWPAUSEDENYControlled execution
Integrations honor the decision
Evidence record
W3C PROV-O provenance
Pause loops back: fetch fresher evidence, clarify or request approval, then re-evaluate.
Every decision is stored with its evidence and policy version as queryable knowledge.
The LLM proposes. TrustGraph decides whether it may act.
04In practice
One business action: a supplier payment
The agent holds credentials for the payment API. Business policy also requires:
- An active supplier relationship
- A valid contract covering the invoice
- A verified beneficiary account
- An amount within delegated authority
- Extra approval above a set threshold
Policy evaluationpolicy v14
SupplierContractInvoiceBeneficiaryApprover
PROPOSEDPayment
ALLOWcompliant
PAUSEneeds approval
DENYprohibited
The result: the agent cannot execute the payment when the rules are unsatisfied.
05The platform
From business understanding to controlled action
Understand
Build an ontology-grounded graph from documents and systems, and connect each action to the business entities, relationships and policies it touches.
Govern
Apply explicit rules across authority, scope, risk, uncertainty, time and value before execution, not after the incident.
Verify
Preserve the evidence and policy version behind every decision, so any outcome can be reproduced and explained.
06Open source · Open standards
Run it anywhere. Own your intelligence.
Built on W3C RDF 1.2, OWL, SHACL and PROV-O, so your ontologies, policies and provenance stay portable. Deploy with Docker Compose or Kubernetes, on every major cloud or on-prem, with hosted LLM APIs or open-weight models on your own accelerators.
- Hosted LLMs
- AnthropicOpenAIGoogleMistral
- Open models
- vLLMLM Studiollama.cppOllamaon NVIDIA, AMD or Intel acceleration
- Clouds
- AWSAzureGoogle CloudOVHcloudScalewayand on-prem
- Deploy
- Docker ComposeKubernetes
Built-in inference stack. Run open-weight models on TrustGraph's own LLM inferencing stack, alongside the context graph and agent runtime, so the whole system can operate inside your environment. The AI factory
$ npx @trustgraph/config ? LLM provider Anthropic ? Deploy target Docker Compose ✓ Generated docker-compose.yaml $ docker compose up -d ✓ TrustGraph is running
07Why teams switch
What teams replace with TrustGraph
Operating graphvs. property graphs
No native ontology enforcement, provenance has to be hand-modeled, and there's no policy runtime for agent actions.
Graph vs. context graphW3C ontologiesvs. closed ontologies
Proprietary ontologies lock business logic inside one vendor. Enterprises want to own their intelligence.
Why open ontologiesDefensible agent executionvs. agent frameworks
Agent frameworks orchestrate model calls and leave the outcome to the LLM. TrustGraph decides every action deterministically against policy and records the evidence, so each outcome can be explained and defended.
What is determinism for AI?Open source, W3C standards and enforcement at the point of action, deployable anywhere.
08Use cases
Built for teams where actions carry real risk
Finance and banking
Payments, approvals and delegated authority enforced on every agent action.
ExploreAgentic threat intelligence
Unify threat, event and log intelligence in one governed graph.
ExplorePharmacovigilance
Traceable drug-safety workflows with evidence behind every decision.
ExploreAerospace, defense and telecom
High-assurance agents where decisions must be auditable and reproducible.
ExploreWhat is TrustGraph?
TrustGraph is an open-source context graph platform with a deterministic agent runtime. It builds an ontology-grounded knowledge graph of how a business operates, then checks every action an AI agent proposes against that graph, returning allow, pause or deny with recorded evidence.
What is a context graph?
A context graph connects business entities, relationships, policies and evidence so AI systems act with the full meaning of a request. Thoughtworks uses TrustGraph as its definition of a context graph.
How is this different from AI guardrails?
Guardrails filter what a model says. TrustGraph governs what an agent does: it resolves the business entities an action affects and evaluates authority, scope, risk, uncertainty, time and value before execution.
Which standards does TrustGraph use?
W3C RDF 1.2, OWL, SHACL and PROV-O. Your ontologies, policies and decision records stay portable with no lock-in.
Is TrustGraph open source?
Yes, the core is open source. Enterprise add-ons cover knowledge-based access controls and IAM, an attestation engine, temporal knowledge management and the deterministic agent runtime, with engineering support for production deployments.
10See it in action
Put your business rules between agents and action.
See a governed agent block, pause and resume a real workflow in 30 minutes.