$ termdex

Coding agent (CLI) · SWE-agent team

mini-SWE-agent

Deliberately tiny research agent: roughly 100 lines of Python for the agent class, bash as its only tool, linear history. Built by the SWE-bench authors as a benchmark baseline, also usable from the command line.

Reviewed by the Termdex team · facts checked 2026-10-28 against 5 vendor sources · how we check

Official site

mini-SWE-agent review

mini-SWE-agent is a deliberately tiny coding agent, written in Python and usable from the command line. It is free and MIT licensed. We read the vendor's pages and repository in October 2026. We did not run an independent test.

Who it is for

Researchers and tinkerers who want a small, transparent agent loop for benchmarking models or building their own harness.

What stands out

The SWE-bench authors at Princeton and Stanford built it as a benchmark baseline. The agent class is roughly 100 lines of Python. Bash is its only tool, and the message history is linear. Each action runs as an independent subprocess, so swapping in `docker exec` or another sandbox is simple.

It needs no tool-calling support from the model. It works through litellm, OpenRouter, Portkey and similar routers, and the docs include a worked vLLM example for local models. Install with `pip install mini-swe-agent`, `uvx mini-swe-agent` or `pipx install mini-swe-agent`.

Price and licence

The LICENSE.md in the repository is MIT, so this is open source. The tool has no pricing, seen 28 October 2026. Model cost comes from the provider you connect with your own keys, or from a local server such as vLLM.

The README reports above 74% on SWE-bench Verified. That is the vendor's claim, and we did not check it.

Where it falls short

Alternatives

See terminal coding agents. For a full-featured open-source agent, try Aider or OpenCode.

Facts: platforms, price and licence

platforms
Python package (pip/uv/pipx); docs do not state supported operating systems. Sandboxing options include Docker/Podman, Singularity/Apptainer and bubblewrap. v2 is current; v1 lives on a separate branch.
price
FreeFree MIT-licensed tool (seen 2026-10-28). No product pricing. Model cost comes from whichever provider you plug in via litellm/OpenRouter/Portkey with your own keys, or from a local server such as vLLM.
source
Open sourceLicence: MIT
github
SWE-agent/mini-swe-agent 8.9k stars, seen 2026-10-28

mini-SWE-agent features

✓ documented · – absent or not in its feature list · ? not verified

mini-SWE-agent strengths

  • The core agent is roughly 100 lines of Python with a linear message history, so it is easy to read, debug and fine-tune against
  • Each action runs as an independent subprocess, so swapping in docker exec or other sandboxes is simple
  • Model-agnostic: needs no tool-calling support, only bash, and works through litellm, OpenRouter, Portkey or local servers
  • Actively maintained by the team behind SWE-bench; README reports above 74% on SWE-bench Verified (vendor claim) and use as a benchmark harness

mini-SWE-agent weak points

  • Minimal by design: bash only, no built-in editing tools, MCP or subagent system documented
  • Research and benchmark focus, so it lacks the editor integration, plan modes and polish of full coding assistants
  • Docs state no supported operating systems; local-model guidance is thin beyond vLLM (Ollama appears only as a config value, LM Studio is not mentioned)
  • Version 2 changed the interface (migration guide exists), so older tutorials may not match

Who mini-SWE-agent is for

best for
Researchers and tinkerers who want a small, transparent agent loop for benchmarking models or hacking on their own harness.
look elsewhere if
Teams wanting an out-of-the-box product with IDE integration, approvals UI and managed support.

Lists that include mini-SWE-agent

Sources and what we could not verify

Checked 2026-10-28. Prices and features change; the vendor's page is the authority.

Not verified