Mini-SWE-agent

tom@tomrochette.com (Tom Rochette) — Sat, 10 Oct 2026 00:00:00 +0000

Mini-SWE-agent is the SWE-bench team’s deliberately minimal coding agent, a roughly hundred-line Python loop around a bash shell that its own parent project now points new users to.

What it is
#

A terminal CLI (mini), a batch SWE-bench runner, a trajectory inspector, and Python bindings, all MIT-licensed and installed from PyPI as mini-swe-agent. It is built by the Princeton and Stanford team behind SWE-bench and SWE-agent, and its design point is subtraction: the model gets a shell, a fixed system template, and nothing else, so the harness disappears from the measurement. Models arrive through litellm, OpenRouter, or Portkey, and execution runs locally or inside docker/podman, singularity/apptainer, bubblewrap, Contree, or Modal environments. This is version 2, a rewrite documented in a migration guide from the v1 that first made the hundred-line claim.

Status
#

Active and heavily used, mostly out of sight: 8,355 stars, pushed 2026-10-06, on a repository created 2025-06-28 (GitHub API, as of 2026-10-10), with v2.4.6 (2026-07-23) the latest release. PyPI counted 11,605,679 downloads in the last month as of 2026-10-10, which is batch-evaluation traffic, not desktop installs. The parent SWE-agent repository (20,514 stars, last release v1.1.0 in May 2025, default branch last pushed 2026-07-16) carries a README notice saying most development effort has moved to mini-swe-agent and recommending it going forward, so the lineage has formally superseded its ancestor.

The team claims over 74 percent on SWE-bench Verified with Gemini 3 Pro, and a self-reported adoption list spanning Meta, NVIDIA, IBM, Anyscale, Essential AI, and Nebius. The independent-footprint signal is thin in the other direction: its Show HN drew 7 points in July 2025, so the adoption argument rests on the download counter and the team’s name, not community discussion.

Strengths
#

The whole harness fits in your head, which is exactly what a baseline needs.
The strongest batch-evaluation surface in this section: SWE-bench runs, trajectory files, an inspector, and Python bindings for subclassing the loop.
Broad model reach through litellm, OpenRouter, and Portkey, including local servers through litellm’s api_base.

Cautions
#

Nothing else ships: no subagents, no MCP, no IDE integration, no hooks, no rules files, because the point is the baseline, not the daily driver.
The hundred-line pitch has drifted: the v2 core agent file is 190 lines, and the package around it is far larger than the marketing line suggests.
The release train has been quiet since v2.4.6 on July 23, 2026, and the adoption list is self-reported.

Pricing
#

Free and MIT-licensed via PyPI. There is no paid tier; you bring your own provider API keys, so the only cost is the tokens your provider charges.

Compared to
#

Pi: the other deliberate-minimalism bet, but a polished daily-driver TUI with extensions, where mini is a research loop you are meant to read and rewrite.
Claude Code: the platform benchmark this project measures against, closed and feature-complete where mini is open and bare.
SWE-agent (no note): the heavier ancestor whose config-driven search and interface mini deliberately deletes; use it only when you need its tool scaffolds rather than the shell.

Bottom line
#

Recommended for researchers who need a readable baseline harness and for teams running batch evaluations. Not for daily driving, and not for anyone who needs subagents, MCP, or editor integration, because none of it exists here by design.

Changes
#

2026-10-10 - Created from the awesome-ai-coding-tools scan, admitted as the live carrier of the SWE-agent lineage that its parent repository now supersedes.

References
#

https://mini-swe-agent.com/latest/ - the overview: hundred-line pitch, v2, adoption claims, and the over-74-percent SWE-bench Verified claim
https://mini-swe-agent.com/latest/faq/ - model support through litellm and OpenRouter, API key handling, and the FAQ frame for mini versus SWE-agent
https://mini-swe-agent.com/latest/models/local_models/ - local models through litellm api_base with the vLLM and Ollama examples
https://api.github.com/repos/SWE-agent/mini-swe-agent - 8,355 stars, MIT, created 2025-06-28, pushed 2026-10-06, as of 2026-10-10
https://api.github.com/repos/SWE-agent/mini-swe-agent/contents/src/minisweagent/agents/default.py - the v2 core agent file at 190 lines, grounding the pitch-drift caution
https://api.github.com/repos/SWE-agent/SWE-agent - the ancestor’s 20,514 stars and v1.1.0 last release
https://api.github.com/repos/SWE-agent/SWE-agent/readme - the supersession notice pointing development effort to mini-swe-agent
https://pypistats.org/api/packages/mini-swe-agent/recent - 11,605,679 downloads in the last month as of 2026-10-10
https://hn.algolia.com/api/v1/items/44682897 - the 7-point Show HN, cited as the thin independent-footprint signal

swe-bench on tomrochette.com

Mini-SWE-agent

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

swe-bench on tomrochette.com

Mini-SWE-agent

What it is #

Status #

Strengths #

Cautions #

Pricing #

Compared to #

Bottom line #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#