diff --git a/agents/SWE-agent__mini-swe-agent/README.md b/agents/SWE-agent__mini-swe-agent/README.md
new file mode 100644
index 0000000..ff02b83
--- /dev/null
+++ b/agents/SWE-agent__mini-swe-agent/README.md
@@ -0,0 +1,63 @@
+# mini-swe-agent
+
+**The minimal AI software engineering agent.** Built by the Princeton & Stanford team behind [SWE-bench](https://github.com/swe-bench/SWE-bench) and [SWE-agent](https://github.com/SWE-agent/SWE-agent).
+
+## What it does
+
+`mini-swe-agent` solves GitHub issues and software engineering tasks by interacting with a computer **entirely through bash** — no custom tool implementations, no special interfaces, no complex scaffolding. Just the model, a shell, and a task.
+
+Scores **>74% on SWE-bench Verified** with a ~100-line Python agent class.
+
+## Key capabilities
+
+- 🐚 **Bash-only** — no tool-calling interface required; works with any chat-completion model
+- 🧩 **Model-agnostic** — runs on Claude, GPT-4, Gemini, and more via [litellm](https://github.com/BerriAI/litellm), OpenRouter, Portkey
+- 📦 **Sandboxed** — local, Docker/Podman, Singularity/Apptainer, bubblewrap, contree
+- 🔬 **Research-grade** — linear trajectory = perfect fine-tuning/RL data
+- ⚡ **Fast startup** — starts much faster than heavier agent frameworks
+- 🏢 **Production-adopted** — used by Meta, NVIDIA, Essential AI, IBM, Princeton, Stanford
+
+## How it works
+
+1. The agent receives a task (a GitHub issue or CLI prompt)
+2. It reasons aloud in a THOUGHT block, then issues **one bash command**
+3. The environment runs the command in a **fresh subshell** (stateless, safe)
+4. The agent observes the output and iterates until the task is solved
+
+```python
+from minisweagent.agents.default import DefaultAgent
+from minisweagent.models.litellm_model import LitellmModel
+from minisweagent.environments.local import LocalEnvironment
+
+agent = DefaultAgent(
+    LitellmModel(model_name="claude-sonnet-4-5-20250929"),
+    LocalEnvironment(),
+)
+agent.run("Write a sudoku game")
+```
+
+## Installation & usage
+
+```bash
+# Quickest — no install needed
+pip install uv && uvx mini-swe-agent
+
+# Or install permanently
+pip install mini-swe-agent
+mini  # run the CLI
+```
+
+## Benchmarks
+
+| Model | SWE-bench Verified |
+|---|---|
+| Gemini 3 Pro | 74%+ |
+| Claude Sonnet 4 | ~70% |
+| GPT-5 | ~68% |
+
+## Links
+
+- 🌐 [Documentation](https://mini-swe-agent.com/latest/)
+- 📄 [NeurIPS 2024 Paper](https://arxiv.org/abs/2405.15793)
+- 🏆 [SWE-bench Leaderboard](https://www.swebench.com/)
+- 💬 [Slack Community](https://join.slack.com/t/swe-bench/shared_invite/zt-36pj9bu5s-o3_yXPZbaH2wVnxnss1EkQ)
diff --git a/agents/SWE-agent__mini-swe-agent/metadata.json b/agents/SWE-agent__mini-swe-agent/metadata.json
new file mode 100644
index 0000000..92d044e
--- /dev/null
+++ b/agents/SWE-agent__mini-swe-agent/metadata.json
@@ -0,0 +1,15 @@
+{
+  "name": "mini-swe-agent",
+  "author": "SWE-agent",
+  "description": "A radically minimal AI software engineering agent solving GitHub issues and CLI tasks via bash only. Scores >74% on SWE-bench Verified with ~100 lines of Python.",
+  "repository": "https://github.com/SWE-agent/mini-swe-agent",
+  "path": "",
+  "version": "2.0.0",
+  "category": "developer-tools",
+  "tags": ["software-engineering", "bash", "github-issues", "swe-bench", "minimal", "litellm", "docker", "coding-agent", "research", "benchmark"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-5-20250929",
+  "adapters": ["system-prompt"],
+  "icon": false,
+  "banner": false
+}