mini-swe-agent

Minimal open-source bash-only coding agent from the SWE-bench team, used as a CLI, Python library and benchmark baseline.

Where it ranks

–
Coding
Coding agents · Local · no rated model available here yet
With cloud models: tier C (19.1% on Terminal-Bench 4.0)

Ranked by its own benchmark result where it has one; otherwise by the best model it offers.

Vendor: SWE-agent team (Princeton & Stanford)
Approach: Coding agents
Where it runs: Runs fully on your hardware with local models; cloud models optional
Interface: Command line, API
Skill level: Advanced: command line, node graphs or self-hosting
Price: Free, open source
Pricing notes: Open source; you pay your own model API costs (any LiteLLM/OpenRouter/Portkey model) or use local models.
Source code: Open source (MIT)
GitHub stars: 8.7k (19 Oct 2026)
Last verified: 19 Oct 2026

Price from mini-swe-agent.com, 19 Oct 2026.

Benchmark	Result	Entry on the leaderboard	As of	Source
Terminal-Bench 2.0	42.5%	`Mini-SWE-Agent + Claude Sonnet 4.5` rank 82 of 142 entries; raw accuracy 42.53 ± 2.81; entry date 2025-09-29; verified; read via the site's leaderboard-read API (page is client-rendered)	19 Oct 2026	tbench.ai
Terminal-Bench 2.1	76.2%	`mini-SWE-agent + Muse Spark 1.1` reasoning effort: xhigh; rank 13 of 22 entries; raw accuracy 76.18 ± 2.41; entry date 2026-07-09; read via the site's leaderboard-read API (page is client-rendered)	19 Oct 2026	tbench.ai
Terminal-Bench 3.0	42.7%	`mini-SWE-agent + Opus 5` reasoning effort: max; rank 1 of 12 entries; raw accuracy 42.70 ± 3.10 (95% CI half-width); entry date 2026-07-24; agent org: Princeton; read via the site's leaderboard-read API (page is client-rendered)	19 Oct 2026	tbench.ai
Terminal-Bench 4.0	19.1%	`mini-SWE-agent + Gemini 3.8 Flash` reasoning effort: high; rank 29 of 35 entries; raw accuracy 19.09 ± 3.36; entry date 2026-09-02; read via the site's leaderboard-read API (page is client-rendered)	19 Oct 2026	tbench.ai