OpenRath#

A PyTorch-like multi-agent and multi-session runtime.

OpenRath turns agent runtime state into explicit Python objects: Session carries collaboration state, Sandbox decides where tools run, Memory persists knowledge across runs, and Workflow composes agents into reusable systems.

Start Tutorials Read Developer Notes GitHub

import os

from rath import flow
from rath.session import Session

agent = flow.Agent(
    "Use tools when helpful.",
    model=os.environ.get("OPENAI_DEFAULT_MODEL") or "gpt-5.5",
    memory="local",
)

user = Session.from_user_message(
    "Create a file, then read it back."
).to("local", spec="./")

out = agent(user)

Tutorials use deterministic steps where reproducibility matters. Production agent workflows use the same Session, FlowToolCall, Workflow, Memory, and Backend abstractions with OpenAI-compatible or Anthropic configuration.

Where To Start#

Path	Use it for	Entry
Installation	Install OpenRath, configure model credentials, and connect a sandbox backend when needed.	Installation
Project Status	Understand v1.2.0 maturity, new runtime layers, and remaining productization work.	Project Status
Tutorials	Learn from the v1.2 numbered example ladder and focused runtime tutorials.	Tutorials
Blog	Read project updates, release announcements, and engineering notes.	Blog
Developer Notes	Understand runtime components, call boundaries, memory, and async behavior.	Developer Notes
API Reference	Look up public modules, function signatures, and integration points.	API Reference

Core Model#

Concept	Role
`Session`	Carries ordered chunk rows, sandbox placement, and lineage metadata through a run.
`Backend`	Opens the local or OpenSandbox execution environment attached to a session.
`Memory`	Persists recalled and committed agent knowledge through local or optional OpenViking stores.
`FlowToolCall`	Exposes JSON schemas to the model and Python callables to the runtime.
`Workflow`	Composes agents and session transformations as ordinary Python modules.
`AgentParam`	Stores the agent system session plus LLM routing options.
`Provider`	Stores LLM routing, request parameters, retry/budget policy, and optional config lookup.

OpenRath in the PyTorch lens — OpenRath borrows PyTorch’s compositional vocabulary for agent runtime state: Session as the flowing value, Backend as placement, Memory as persistent state, Tool as callable operation, and Workflow as composition.#

OpenRath paradigm map — The runtime separates agent composition, session state, backend placement, and memory into explicit surfaces.#

Runnable Example Ladder#

Example	Demonstrates
01 Hello Agent	The smallest `flow.Agent` program.
02 Session Lineage	Fork, detach, session graph, and JSONL export.
09 Memory	Key-free local memory remember / recall / commit.
10 Provider Variation	Swapping OpenAI-compatible and Anthropic providers.

PyTorch Mental Model#

PyTorch mental model	OpenRath counterpart
Tensor carries data	`Session` carries agent state
Module composes computation	`Workflow` / `Agent` composes behavior
device controls placement	`Backend` controls execution placement
Parameter persists state	`Memory` persists agent knowledge
callable module exposes a reusable interface	`FlowToolCall` exposes tools