Report Abuse

Basic Information

SWE-ReX is a runtime interface designed for developers building AI agents that need to execute shell commands in sandboxed environments. It provides a consistent API so agent code can run commands locally or remotely without changing logic. The repository targets scenarios where agents must interact with real shells or command line tools, including interactive programs, and where runs must scale to many parallel sessions. SWE-ReX was created from the SWE-agent project to simplify infrastructure concerns and to make it easy to run agents across Docker containers, cloud machines, Modal, and other execution backends. It focuses on enabling reproducible, massively parallel agent evaluation and execution while keeping agent implementation independent from the underlying environment. The project includes packaging and optional extras for different backends and points users to documentation for detailed setup and usage.

Links

Categorization

App Details

Features
SWE-ReX provides a set of practical features for agent-driven remote execution. It recognizes when shell commands finish and captures output and exit codes for the agent. It supports interactive command line tools so agents can use programs like ipython or gdb in a session. The framework can manage multiple shell sessions concurrently, allowing agents to operate multiple terminals or REPLs at once. It is backend-agnostic and works with local execution, Docker containers, AWS remote machines, Modal and other platforms, and it is designed to support non-Linux hosts. The project offers optional pip extras for Modal, Fargate, Daytona (WIP) and a development bundle. It includes documentation and packaging to help integrate the runtime into existing agent workflows.
Use Cases
SWE-ReX helps teams focus on agent logic and evaluation instead of infrastructure glue. By disentangling execution backends from agent code, experiments and benchmarks become easier to run and reproduce. The framework supports fast, massively parallel runs so you can evaluate many agents or tasks in parallel, with the README explicitly noting large-scale runs such as 30 SWE-bench instances and the ability to run dozens or hundreds of agents concurrently. Its broad platform support reduces friction when moving between local, containerized, and cloud environments and improves stability of agent experiments. The tooling is aimed at simplifying agent development, enabling interactive debugging within sessions, and accelerating large benchmark evaluations powered by SWE-agent.

Please fill the required fields*