Build. Test. Launch.
RepoLaunch Agent:
Turning Any Codebase into Testable Sandbox Environment
Meet RepoLaunch, an agent that automatically builds and manages code repositories across programming languages and operating systems. From source code to a ready-to-test sandbox.
What RepoLaunch Can Do
From repository to reproducibilityRepoLaunch is the first agentic SWE tool for repository build and test management across programming languages and operating systems. Give it a repository, and it handles the groundwork.
Ready-to-run environments
Install dependencies and build the repository in a Docker image, with layer information for Dockerfile reconstruction.
Make changes, build again.
Get organized commands to rebuild your repository inside its container after source code changes.
Structured, traceable testing.
Get test commands, parse output into structured testcase--status mapping, optionally find commands to run individual testcases.
Applications
Explore the benchmark ↗RepoLaunch powers the build and test of the execution environments behind the whole SWE-bench-Live Software Engineering dataset family. RepoLaunch has been used by many organizations to benchmark code language models and coding agents, and for agentic training of Language Models through rejection-based SFT and reinforcement learning. The SWE-bench-Live family RepoLaunch creates include:
Latest News from RepoLaunch
Project news & updates- A little memory goes a long way.
A memory-aware solution reuses successful RepoLaunch results across commits of the same repository. Across 856 GitHub issues from 93 repositories, experiments show ≥98% success, 82% savings on language model API cost, and 78% savings on Docker image storage. Particularly useful for creating multiple task instances from a repository’s issues. Read the development guide ↗
- More models. More possibilities.
LiteLLM support brings compatibility with mainstream LLM providers and local deployments for agentic training (RFT and RL) based on launch results. RepoLaunch retains the plain-text Thought–Action format for compatibility with smaller open-source models that struggle with tool-call fields.
- Building environments for agentic RL.
Thanks to the GLM-5 Foundation Model team for using RepoLaunch to create executable environments for agentic reinforcement learning.