Our flagship project, OpenRath, is a PyTorch-like framework for dynamic multi-agent and multi-session workflows.

OpenRath treats agent runtime state as explicit, composable Python objects:

• Session carries conversation state, branches, lineage, and usage.
• Sandbox decides where tools execute.
• Memory persists knowledge across runs.
• Tool exposes callable operations to models.
• Agent transforms one session into another.
• Workflow composes agents and workflows into larger systems.
• Selector routes between workflows while control flow stays plain Python.

This structure keeps state, execution, memory, and orchestration separate while connecting them through a single flowing value: Session.

pip install openrath

Then explore the examples or read the documentation.

• OpenRath — the open-source runtime for multi-agent and multi-session systems.
• OpenRath Documentation — guides, concepts, API reference, and runnable examples.

Questions, ideas, issues, and pull requests are welcome.

Open an issue · Contribute