unclecheng-li/vulnclaw

This page was translated automatically. Technical details may differ from the original. View the original

3.8K stars 534 forks Quality 73 Trend 57MIT

Based on AI Agent + MCP toolchain + penetration skill orchestration, combined with large language models, natural language input→ automatically completes the entire process of "information gathering→ vulnerability discovery → vulnerability exploitation → report generation."

Overview

Project Official Website: https://unclecheng-li.github.io/vulnclaw.com/ Based on LLM Agent + MCP toolchain + penetration skill orchestration, combined with compatible models such as OpenAI, Anthropic, MiniMax, and DeepSeek, natural language input → automatically completes the entire process of "information gathering→ vulnerability discovery, → vulnerability exploitation → report generation." - — Abandoning fixed-turn workflows, using "goal achieved / exploration frontier exhaustion / security budget" as termination conditions, automatic convergence — modeling penetration as a search from origin to goal: Fact (confirmed fact) + Intent (exploration direction), structurally avoiding "going in circles" — Claimed flags/conclusions must appear character by character in real tool output to be accepted, eliminating false victories by fabricating flags — Describe penetration intent in plain language, Automatic recognition stages and tools — OpenAI / Anthropic / MiniMax / DeepSeek / Zhipu / Moonshot / Qianwen / SiliconFlow / Doubao / Baichuan / Jieyue Xingchen / SenseTime / Zero One Everything, one-click switching — 4 MCP services: fetch / memory native out-of-the-box, chrome-devtools / burp integration with external MCP services for browser automation and HTTP packet capture and playback — — VulnClaw's proprietary packet capture: filtered by in-run scope, then appended JSONL indexes + raw packets are placed on evidence/traffic/ for each request. Built-in traffic_list / traffic_view / traffic_repeat / traffic_sitemap tools can directly read and write the storage (traffic_repeat supports overwrite replay), and reports of verified vulnerabilities are directly inscribed to prove their original request/response. The mitmproxy proxy and Playwright browser capture backend as optional dependencies (pip install vulnclaw[traffic]) and enabled according to availability checks; It is automatically accessed in sandbox loops with the sandbox/runtime directory PRD implementation.

Overview (machine translated)

*AI-driven penetration testing CLI tool — speak plain language, exploit vulnerabilities. *

🌐 English version: README_EN.md

**This project is an independently runnable AI penetration testing agent. **

Project Official Website: https://unclecheng-li.github.io/vulnclaw.com/

Based on LLM Agent + MCP toolchain + optional Skill reference materials, Compatible with models such as OpenAI, Anthropic, MiniMax, and DeepSeek, Natural language input → automatically completes the entire process of “information gathering → vulnerability discovery → vulnerability exploitation → report generation.”

Quick Start · Architecture · Built-in Skills

What can it do?

Input natural language, AI automatically executes the entire penetration testing process:

Suitable for authorized penetration testing, CTF competitions, safety training, red team drills, and other scenarios.

Features

Model-Led Solution Engine (default) — Autonomous loops similar to Claude Code/Codex, where the model decides the next step, when to call tools, and when to complete/query/determine with no way forward
AgentState Evidence Memory — Tool results are uniformly written AgentState.evidence, raw original text is fully retained; Active context by default only injects high-signal previews, evidence_search / evidence_view are used to revisit original evidence on demand
Lightweight Correction Layer — Records signals such as repeated calls, failed downgrades, time-consuming, and new discoveries before and after tool calls; Repeated reading of the same evidence range is suppressed, continuous evidence idling triggers stall guard, but the old stage planner is not restored
Evidence-Level Anti-Illusion Gate — Claimed flags/conclusions must appear character by character in real tool outputs to be accepted, eliminating false victories by fabricating flags
Natural Language Driven — Describe penetration intent in human language, automatically identify stages and tools
13 LLM Providers — OpenAI / Anthropic / MiniMax / DeepSeek / Zhipu / Moonshot / Qianwen / SiliconFlow / Doubao / Jieyuexing Chen / SenseTime / OpenRouter / Local Ollama, one-click switching
MCP Toolchain — 4 MCP services: fetch / memory local out-of-the-box, chrome-devtools / burp integration with external MCP services for browser automation and HTTP packet capture and playback

This page was translated automatically. Technical details may differ from the original.

Install

docker run --rm -it \

New to MCP? Start with the mcpskill.org guide →