FishCodeTech/muteki

Project Muteki – A Multi‑Agent Security Automation Framework

What it is

• An open‑source, Python‑centric framework that lets you run a swarm of large‑language‑model (LLM) agents to solve penetration‑testing and CTF challenges automatically.
• A Coordinator (the “Planner”) schedules a set of Worker engines – Claude, Codex, Cursor, Pi, OMP, Kimi, Grok and OpenCode – each of which talks to a vendor‑specific CLI, runs commands, and returns evidence.
• The UI is a web app (backend on port 8000, frontend on 3001) where you configure agents, upload challenge data, and watch the swarm’s progress.

Key capabilities

• Fully‑automated CTF solving – the authors report 200/200 on the NYU CTF Bench, top‑10 placement at RIFFHACK 2026, and first‑blood on the Yunjing blackmaze range.
• Hybrid runtime – Workers can run locally on the host machine or inside Docker containers; the latter ships a pre‑built image with a full CTF toolchain.
• Planner & Titler models – a separate LLM endpoint decides which steps to take and can generate titles for each step.
• Evidence tracking – every command, tool call and candidate flag is stored in a per‑run workspace, and a markdown write‑up can be generated after a run.
• Multiple workspaces – besides the default single‑task mode, there are experimental Conversation and Competition workspaces for multi‑turn dialogue and remote CTF platform integration.

How to get started

1. Install the prerequisites – Python ≥ 3.13, uv, Node.js/npm, and at least one supported Agent CLI (e.g., Claude, Codex, etc.).
2. Clone the repo and run the provided init.sh script to install Python dependencies.
3. Start the web UI with ./run.sh web and open http://127.0.0.1:3001.
4. In the UI, add your Agent credentials (API keys or CLI logins), enable one or more Workers, choose Local or Container runtime, and configure a Planner endpoint.
5. Create a single‑task by entering the challenge description, target URL, flag format, and any files, then click Dispatch swarm.
6. Watch the Coordinator’s messages, Workers’ tool calls, and candidate flags appear in real time. You can pause, freeze, or stop the swarm at any point.

Typical workflow

Phase	Action
Setup	Install OS‑specific scripts (./ctf-tools/setup.sh on macOS, setup-ubuntu.sh on Linux) to get common pentest tools (nmap, gobuster, etc.).
Configure	Add API keys/tokens for the chosen LLM engines, enable Workers, pick a Planner model.
Run	Submit the challenge; the Planner creates Steps, Workers execute commands, and evidence is collected automatically.
Review	Examine the generated evidence, copy any candidate flag, mark false positives, or ask the system to generate a write‑up.
Iterate	Restart the swarm with the existing workspace to continue hunting for additional flags.

Safety & legal notice

• Muteki can execute arbitrary commands on target services, so the authors stress running it only on systems you own or have explicit permission to test.
• The repository includes a disclaimer that the author is not responsible for any unauthorized attacks performed with the tool.

Release status

• Latest stable version 0.4.5 (May 2026). Includes a macOS desktop app (Apple Silicon & Intel) and support for remote Muteki services. Windows/Linux desktop bundles are not shipped yet.
• Eight active CLI engines; DeepSeek Harness is listed but currently unavailable.
• The project is actively maintained (star‑history chart, issue tracker, discussion QR code).

Who might use it

• Security researchers and CTF teams looking to automate repetitive exploitation steps.
• Developers interested in building multi‑LLM agent swarms for security‑oriented tasks.
• Anyone experimenting with “AI‑driven penetration testing” who wants a ready‑made orchestration layer.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.