Open infrastructure for durable multi-agent and multi-session systems.
We build agent runtimes that are explicit in Python, recoverable in production, and inspectable from workflow definition to final artifact.
Website · Documentation · OpenRath · Examples · PyPI · Paper
English | 简体中文
OpenRath is a durable, PyTorch-like runtime for multi-agent and multi-session systems.
OpenRath v2.0.0 marks OpenRath's move into production environments. It compiles
explicit @step and @router workflow boundaries into immutable execution
plans, then persists Runs, Events, Checkpoints, Interrupts, and effect outcomes
across worker restarts.
@step / @router
|
v
ExecutionPlan -> Run -> Event -> Checkpoint
| |
| +-> Interrupt / Effect ledger
|
+-> embedded SQLite
+-> PostgreSQL + optional Redis + S3-compatible artifacts
|
v
Agent Server HTTP + SSE
The v1 Python façade remains available: Session carries agent state,
Sandbox controls execution placement, Memory persists knowledge, Tool
exposes operations, and Agent and Workflow compose behavior.
| Repository | Purpose | Start here |
|---|---|---|
| OpenRath | Runtime, Python SDK, Agent Server, deployment references, tests, and release evidence | pip install openrath |
| OpenRath-Docs | Version-matched concepts, tutorials, API reference, migration, security, and operations guidance | docs.openrath.com |
| OpenRath-Example | Runnable Colab workflows for literature research, PowerPoint creation, and Excel visualization | Open in Colab |
The framework repository also contains a 12-step core API ladder. Use it to learn one concept at a time; use OpenRath-Example to inspect complete durable workflows and their final deliverables.
- Explicit state. Workflows, sessions, plans, revisions, Runs, and effects are inspectable objects instead of hidden loop state.
- Durable recovery. Checkpoints, leases, fencing, retries, cancellation, and Interrupts make failures and human decisions part of the runtime model.
- Safe boundaries. Authentication, action grants, tenant scope, policy, secret references, trust labels, and audit are separate contracts.
- Composable infrastructure. Provider, Tool/MCP, Sandbox, Memory, storage, artifact, telemetry, and evaluation systems plug into governed interfaces.
- Evidence-driven releases. Immutable artifacts, migrations, tests, SBOMs, scans, benchmarks, soak reports, and recovery drills are tied to source revisions.
Install OpenRath:
pip install openrath
Run a durable workflow locally with SQLite, then move to the Agent Server profile with PostgreSQL:
pip install "openrath[server,postgres]"
openrath-migrate
openrath-migrate --checkChoose your next step:
- Learn the core API
- Run an end-to-end Colab workflow
- Read the OpenRath documentation
- Review deployment and operations
- Read the paper
Questions, design ideas, bug reports, and pull requests are welcome.