Kali MCP — 多智能体渗透测试集群

确定性编排的 17-agent 渗透测试集群

中文 | English

中文

定位

Kali MCP 是一个确定性的多智能体渗透测试集群：通过 MCP 协议或命令行，把「对某目标做渗透测试」这一自然语言任务，自动分解为侦察、扫描、利用、验证等子任务，调度 1 个协调器 + 17 个专业智能体并发执行真实工具，最终聚合出带严重性分级的结构化报告。

架构

flowchart TB
    U[自然语言任务] --> IA[IntentAnalyzer 意图分析]
    IA --> TD[TaskDecomposer 任务分解<br/>攻击面测绘/内容发现/漏洞假设/验证固化]
    TD --> SCH[AgentScheduler 波次并发调度<br/>依赖满足即并发，跨阶段保序]
    SCH --> A[17 个专业 Agent]
    A --> RA[ResultAggregator 结果聚合<br/>否定上下文门控，消除 clean 误报]
    RA --> R[结构化报告 + findings]

17 个专业智能体

协调器 CoordinatorAgent 承接 report_generator 与无可用智能体的兜底任务。

核心特性

快速开始

方式一：MCP（自然语言编排）

在支持 MCP 的客户端（Claude Desktop / Claude Code / Oh My Pi）配置：

{
  "mcpServers": {
    "kali": {
      "command": "python",
      "args": ["mcp_server.py", "--tool-profile", "harness"],
      "env": {
        "KALI_MCP_FORCE_ENABLE_MODULES": "multi_agent",
        "K4_LEGACY_CLUSTER": "1"
      }
    }
  }
}

harness 档位暴露 13 个编排工具，核心三入口：

另有 start_task / run_surface_chain（surface playbook 顺序执行）/ board_snapshot / task_* / verify_finding 等任务板与验证工具。

方式二：CLI 实时可视化

# Windows
C:/Windows/py.exe -3 agent_live.py "对 http://localhost:8000/ 做 web 漏洞扫描：目录枚举、CMS识别、注入检测" --no-cache --timeout 300

# Linux / macOS
python3 agent_live.py "对 http://localhost:8000/ 做 web 漏洞扫描" --agents recon,web_vuln --no-cache

agent_live.py 逐行分色打印每个智能体的调度决策与工具调用结果，Windows Terminal 分屏即可获得类 tmux 的实时观察体验。

编排流水线

一条任务经过四阶段确定性编排（以 Web 目标为例）：

1. 攻击面测绘：nmap / masscan / subfinder / amass → recon×2 + subdomain×2

2. 内容发现：gobuster / dirb / ffuf / feroxbuster → web_recon×4

3. 漏洞假设：nuclei_web / nuclei / nikto / wpscan → vuln_scanner×4

4. 验证固化：sqlmap / intelligent_xss_payloads / 命令注入深挖 → web_vuln×4

surface playbook（kali_mcp/core/playbooks/）提供 web_surface / api_surface / auth_surface / svc_surface / internal_lateral / stealth / chain / ai_guided 八类标准打法，run_surface_chain 可顺序执行。

实测数据（2026-08-13 回归）

后端执行

执行后端由 resolve_backend() 启动时动态解析：本地 PATH 有 Kali 工具走 local；检测到 SSH 配置走 ssh 后端（paramiko 免密）；容器环境走 docker。无需硬编码远程地址。

合规声明

本项目仅用于已获书面授权的渗透测试、CTF 竞赛、安全研究与防御性评估。使用前通过 set_engagement_context 声明授权范围；越权扫描、破坏性操作、未授权攻击严格禁止。

License

MIT License — 详见 LICENSE。

Related MCP server: MCP Kali Server

English

Positioning

Kali MCP is a deterministic multi-agent penetration testing cluster: one natural-language instruction is decomposed into recon, scanning, exploitation, and verification subtasks, dispatched across 1 coordinator + 17 specialized agents running real tools concurrently, and aggregated into a severity-graded structured report.

The full pipeline — intent analysis → task decomposition → scheduling → aggregation — is deterministic and runs without any LLM API key; LLM keys only enable optional deep-reasoning enhancement.

Architecture

instruction → IntentAnalyzer → TaskDecomposer → AgentScheduler (wave-concurrent)
           → 17 agents → ResultAggregator (negation-gated) → structured report

17 Specialized Agents

• Information gathering: recon_agent, subdomain_agent, web_recon_agent

• Vulnerability discovery: vuln_scanner_agent, web_vuln_agent, auth_agent, network_vuln_agent, vuln_verifier_agent

• Exploitation: exploit_agent, privilege_agent, lateral_agent

• Specialized: code_analyze_agent, code_audit_agent, crypto_agent, forensics_agent, pwn_agent, source_code_agent

Quick Start

{
  "mcpServers": {
    "kali": {
      "command": "python",
      "args": ["mcp_server.py", "--tool-profile", "harness"],
      "env": {
        "KALI_MCP_FORCE_ENABLE_MODULES": "multi_agent",
        "K4_LEGACY_CLUSTER": "1"
      }
    }
  }
}

Three core MCP entries: agent_run (full pipeline), agent_status (cluster health), kali_run (single-tool fallback).

For real-time visualization:

python3 agent_live.py "web vuln scan on http://localhost:8000/" --no-cache

Measured (2026-08-13 regression)

• Cold run: 174.76s (under 300s cap, wave-concurrency + -c 50)

• Session: completed, 17/17 tasks, 0 failures

• Scheduling success: 100% (16/16)

• Findings: 1 real high-severity SQL injection, XSS clean false-positive eliminated

Disclaimer

For authorized security testing only. Users are responsible for complying with all applicable laws and regulations.

⭐ Star this repo if you find it useful!