OpenRath
An open-source Python runtime for multi-agent workflows, with local tool execution, persistent memory and traceable session branches. BSD-3-Clause licensed.
OpenRath is an open-source Python runtime for developers building AI agent workflows that need to preserve the work behind an answer. Its central object is a session, which keeps conversation context, tool results, execution location and validation evidence together. That gives engineering and research teams a record they can inspect, branch, resume and compare.
Sessions can split into separate paths and merge compatible branches. Their history records which agent, tool call and workspace contributed to a result, including rejected paths. Reusable agents and nested workflows let developers compose larger systems while keeping routing and control flow in Python.
Tool execution can run in a local process or through OpenSandbox. Built-in tools handle files, shell commands and code execution; developers can also add Python tools or connect stdio MCP tools. The runtime records where execution happens alongside the session's other evidence.
Memory persists across runs. The included local backend stores data on the user's machine and can retrieve entries with lexical search without an LLM. It can also use a configured embedding provider, while OpenViking supplies an optional external memory backend.
OpenRath suits workflows where intermediate evidence matters: engineering tasks with diffs and test results, research with sources and experiment state, or trading agents with policy constraints and risk checks. It uses the BSD-3-Clause license.
