Mini SWE-Agent is an open-source AI software engineering agent from the SWE-bench team that solves GitHub issues and assists in your command line with a radically simple design: the core agent is around 100 lines of code. It has earned more than 8,700 stars on GitHub, is actively maintained with updates through October 2026, and scores above 74% on SWE-bench Verified despite its size.
Mini-SWE-Agent is the minimalist sibling of SWE-agent, built on a bet that a coding agent does not need huge configs or a giant monorepo to be effective. It gives the model a bash shell, lets it act, observe, and iterate, and stops there. That simplicity makes it exceptionally hackable: researchers can read the entire agent loop in one sitting, swap models, and instrument every step. It supports multiple model providers through a unified interface, runs locally or in sandboxes, and ships with a command-line interface for everyday use as well as batch runners for benchmark evaluation. The project now powers external SWE-bench harnesses, and the team publishes tutorials on building minimal agents, making it a common starting point for agent research.
Mini SWE-Agent is free and open source under the MIT license. You run it on your own machine and pay only for the model API calls it makes (or nothing beyond compute if you use a local model). There is no paid tier or hosted version sold by the project.
Mini SWE-Agent is for agent researchers who want a transparent baseline they fully understand, developers who want a lightweight CLI coding assistant without heavyweight tooling, and teams evaluating models on real software engineering tasks. If you learn by reading code, this is the coding agent you can actually read.