DuraGraph

v0.2.1 Apache-2.0

What is DuraGraph?

DuraGraph is an open-source control plane for AI workflows — a self-hosted orchestration engine that runs multi-step agent pipelines with full observability, fault tolerance, and auditability.

Define workflows as graphs. Execute them reliably. Observe everything.

Architecture

DuraGraph is built on event sourcing and CQRS — patterns proven in financial systems and distributed databases:

• Event Sourcing — Every state change is an immutable event. Replay any workflow execution from the first token to the last. Full audit trail by default.
• CQRS — Commands write events, queries read projections. Scale reads and writes independently.
• Transactional Outbox — Events are published to NATS JetStream via a transactional outbox, guaranteeing at-least-once delivery with no dual-write problems.
• PostgreSQL — Durable event store and read-model projections. No proprietary state stores.
• NATS JetStream — Reliable event streaming for worker coordination and real-time updates.
• Custom Graph Engine — Supports conditionals, loops, parallel execution, and human-in-the-loop patterns.

This architecture gives you deterministic execution, point-in-time state reconstruction, and zero mutable state.

Who is DuraGraph For?

• AI platform teams building internal orchestration layers with strict reliability requirements.
• Enterprises that need compliance, audit trails, and data sovereignty for AI workflows.
• Teams scaling AI agents that have outgrown simple LLM chains and need production infrastructure.
• Open-source contributors looking for a well-architected orchestration stack to build on.

What’s Included

flowchart LR
  client["Client SDKs<br/>Python · Go · TypeScript"]
  api[API Layer]
  commands[Commands / Queries]
  engine[Graph Engine]
  events[(Event Store)]
  nats[NATS JetStream]
  observability[(Metrics / Tracing)]

  client --> api --> commands --> engine
  engine --> events
  engine --> nats
  nats --> observability

Non-Goals

• Providing a custom LLM runtime (we integrate with existing providers — OpenAI, Anthropic, any HTTP-compatible model).
• Becoming a full-featured cloud platform (our scope is orchestration, not hosting).

LangGraph Compatibility

DuraGraph implements a compatibility layer with LangGraph Cloud APIs. If you have existing LangGraph workflows, see the migration guide for a step-by-step walkthrough.

Success Metrics

We measure success by:

• Reliability: demonstrated ability to recover from failures without data loss.
• Adoption: number of teams running DuraGraph in production.
• Community growth: contributions and RFCs from external developers.