- fix(agents): count billed calls that fail to parse against costlimit
mini-SWE-agent by Princeton University & Stanford University (SWE-agent team)
Overview
Open-source software-engineering agents from Princeton and Stanford researchers: SWE-agent (2024) and its roughly 100-line successor mini-SWE-agent, which uses only bash and works with any model through LiteLLM and similar routers. The SWE-agent README says development effort has moved to mini-SWE-agent.
Status: As of 2026-10-05 the SWE-agent README says it has been superseded by mini-SWE-agent (last SWE-agent release v1.1.0 on 2025-05-22).
mini-SWE-agent was downloaded 3.4M times from PyPI in the week ending Oct 4, 2026, up 58.9% on the week before and down 26.3% over the last 30 days compared with the 30 before. The GitHub repository has 8,225 stars, 1,262 of them added in the last 30 days. The latest stable release is 2.4.6, published Jul 23, 2026. On Terminal-Bench 4.0 its best listed result is 19.1% with Gemini 3.8 Flash, ranking 4 of 4 agents on that board. It was mentioned in 2 Hacker News posts and comments over the last 30 days.
Recent developments
- 2026-02-02: mini-SWE-agent v2 first alpha published on PyPI (2.0.0a1)
mini-SWE-agent usage and attention over time
Download the raw daily series: mini-swe-agent.csv
Install mini-SWE-agent
| PyPI | pip install mini-swe-agent |
|---|
These commands use the packages tracked on this page. The vendor may recommend a different installer; see the official documentation.
Key facts
| Maker | Princeton University & Stanford University (SWE-agent team) |
|---|---|
| Type | Terminal agent (research) |
| First released | April 2024 |
| License | MIT |
| Interfaces | CLI |
| Models | Any model via LiteLLM, OpenRouter, Portkey and similar |
| Local models | Local servers such as vLLM through LiteLLM (docs) |
| Bring your own API key | Yes |
| Main language | Python |
| Open issues and PRs | 88 |
Pricing
Free MIT-licensed research software; users pay their model provider through their own API keys.
, checked Oct 5, 2026. Prices change often; confirm before you buy.
Pulse Score breakdown
Each component is scored 0 to 100 from public signals; a dash means no data for it. Methodology.
What changed: recent mini-SWE-agent releases
1 stable releases in the last 90 days · Full changelog
- fix(config): read config files as UTF-8 regardless of locale
- fix(models): don't mask malformed tool-call errors as truncation
- fix: remove startup-banner emoji that crashes import on non-UTF stdout
- fix(benchmarks): don't mutate shared config across worker threads in getsbenvironment
- @EnesYilmazcode made their first contribution in
- fix(run): pass the startupcommand as a dict to env.execute() function.
- Reject empty config override keys
- @Ethan-Kuang-6666 made their first contribution in
- @CodingFeng101 made their first contribution in
- Update: programbench uses v6 images
- fix: kill local process group on timeout
- @Swing-G made their first contribution in
- maaz139 for reporting the timeout bug
mini-SWE-agent benchmark results
| Benchmark | Model | Effort | Tasks solved | Cost of run | Place |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Gemini 3.8 Flash | high | 19.1% ± 3.4 | $1,829 | #4 of 4 |
| Gemini 3.7 Flash | high | 11.2% ± 2.5 | $1,262 | ||
| Terminal-Bench 3.0 | Opus 5 | max | 42.7% ± 3.1 | $5,818 | #1 of 6 |
| Terminal-Bench 2.1 | Muse Spark 1.1 | xhigh | 76.2% ± 2.4 | $198 | #5 of 6 |
| Terminal-Bench 2.0 | Claude Sonnet 4.5 | — | 42.5% ± 2.8 | — | #40 of 43 |
| GPT-5-Codex | — | 41.4% ± 2.8 | — | ||
| Claude Opus 4.1 | — | 35.1% ± 2.5 | — | ||
| GPT-5 | — | 33.9% ± 2.9 | — |
Up to four best runs per board. Scores are not comparable between Terminal-Bench versions. Source: Terminal-Bench (tbench.ai).
mini-SWE-agent alternatives
| # | Agent | Pulse | Downloads 7d | 7d | 30d | VS Code installs | Stars | Latest release | Price | Last 90 days | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |
OpenAI CodexOpenAI |
82 | 25.6M | up 5.7% | up 21.2% | 15.2M | 128K+153/day | 0.160.1yesterday | Free + $8/mo | ||
| 3 |
Claude CodeAnthropic |
80 | 14.8M | up 11.6% | down 36.4% | 27.1M | 150K+161/day | 2.1.290yesterday | $20/mo | ||
| 5 |
PiEarendil Inc. (created by Mario Zechner) |
75 | 5.2M | up 29.8% | up 52.1% | — | 113K+378/day | 1.0.4yesterday | Free (OSS) | ||
| 7 |
OpenCodeAnomaly (SST team) |
73 | 3.3M | up 22.0% | up 6.1% | 1.2M | 212K+200/day | 1.18.346d ago | Free + $10/mo | ||
| 24 |
Gemini CLIGoogle |
57 | 453K | up 4.7% | up 1.8% | 1.1M | 107K+24/day | 0.62.07d ago | Varies | ||
| 26 |
Oh My PiStencil Labs (Can Bölük) |
56 | 241K | up 11.4% | up 53.9% | — | 34.4K+111/day | 18.6.12d ago | Free (OSS) | ||
| No agents match that filter. | |||||||||||
mini-SWE-agent FAQ
How much does mini-SWE-agent cost?
Free MIT-licensed research software; users pay their model provider through their own API keys.
Is mini-SWE-agent open source?
Yes. mini-SWE-agent is open source, released under the MIT license.
How popular is mini-SWE-agent?
mini-SWE-agent was downloaded 3.4M times from PyPI in the week ending Oct 4, 2026, up 58.9% on the week before and down 26.3% over the last 30 days compared with the 30 before. The GitHub repository has 8,225 stars, 1,262 of them added in the last 30 days. The latest stable release is 2.4.6, published Jul 23, 2026.
What is the latest version of mini-SWE-agent?
The latest stable release we track is 2.4.6, published on Jul 23, 2026.
What are the alternatives to mini-SWE-agent?
The closest alternatives in the same category by Pulse Score are OpenAI Codex, Claude Code, Pi, OpenCode.
Where these numbers come from
PyPI: mini-swe-agent. GitHub: SWE-agent/mini-swe-agent. Site rank: mini-swe-agent.com on the Tranco list. See data sources for how each one is collected.
Product facts and pricing were checked against: github.com, pypi.org, mini-swe-agent.com.
