ARCANGEL0/Mitsu

Meet Mitsu, your chibi red-team operator. Connect your AI provider, set the scope, and she can take an authorized pentest from recon and attack-surface mapping through vulnerability discovery, > exploitation, and post-exploitation, armored with 165+ MCP-connected tools and live session intel. Guide every step or let her pwn autonomously. ( ◠‿◠)

What it solves

Mitsu tackles the heavy lifting in authorized penetration tests. Instead of a human manually running dozens of separate security tools and stitching results together, Mitsu runs as an agentic framework that can take an engagement from initial reconnaissance all the way to post-exploitation on her own. She keeps persistent per-session intelligence (targets, IPs, services, vulnerabilities, findings, accounts, evidence) and uses that context to decide what to check next, saving time and reducing oversight burden.

How it works

Mitsu connects a large language model to a catalog of 165+ red-team tools via an MCP server. The model runs in a tool loop: it reviews the conversation and session Intel, requests a tool when needed, gets the output, updates the Intel, and repeats. Three operating modes control autonomy: Recon (passive, read-only, operator-led), Guided (operator steers, Mitsu pauses for approval before each action), and AutoPwn (Mitsu chooses evidence-led steps autonomously within scope, but still asks for approval on installs and sudo). Each session stores its conversation and Intel in a JSON file, so you can resume, search, rewind, or export. Vuln research combines Intel with search engines and reference databases. The system asks for approval before installing missing tools or requiring credentials.

Who it’s for

Security professionals performing authorized penetration tests, red-team exercises, and vulnerability assessments. It suits both operators who want to direct every step (Guided mode) and those who want a more hands-off autonomous assistant (AutoPwn). It requires Python 3.10+ and a model provider (OpenAI, Anthropic, OpenRouter, Ollama, DeepSeek, Moonshot, NVIDIA NIM, Cloudflare, or custom endpoints). Because it is agentic and can run active checks, it’s intended only for environments you are explicitly authorized to test.

Highlights

  • 165+ integrated tools across 17 categories (OSINT, network mapping, web app testing, exploitation, AD, cloud, forensics, etc.) exposed through MCP.
  • Three modes — Recon, Guided, AutoPwn — to balance autonomy and control.
  • Persistent session Intel per engagement: targets, findings, services, vulnerabilities, domains, IPs, accounts, and observations.
  • Custom tool wrappers and external MCP servers can extend Mitsu’s kit.
  • Multi-provider model support including local Ollama.
  • Interactive TUI with Intel tabs, approval dialogs, and slash commands.
  • Missing tools auto-install (with user approval) via uv.

Related

  • Project
    0sec-labs/0

    0 is an open‑source, AI‑driven security‑workflow platform. Install a tiny binary, configure an LLM provider, then run pre‑built or custom security templates (code review, dependency scan, smart‑contract audit, etc.) via a browser UI, CLI, or remote MCP server. Results are stored locally with evidence and suggested fixes, and the system integrates with common dev‑ops tools (GitHub, CI, Slack, etc.). It’s a research‑preview tool licensed MIT/Apache‑2.0.

  • Project
    soxoj/awesome-osint-mcp-servers

    An awesome-list style compilation of MCP (Model Context Protocol) servers that connect OSINT tools to AI assistants, enabling natural-language queries for investigations across social media, networks, companies, and more.

  • Project
    FluidInference/FluidUse

    FluidUse is a Swift library for on‑device form‑filling and typed‑decision making on macOS. It reads UI fields via Accessibility, queries tiny Core ML models (form‑filler, Laya, GLiNER2, Kev, Clef‑Vision, Intern‑Decision, Decision 2.0, Vela 2.0, Short‑Reply) and types or acts on the answers, all locally with sub‑millisecond latency on Apple Silicon.

  • Project
    Kohaku-Lab/KohakuTerrarium

    KohakuTerrarium is a Python framework for building AI agents. It provides a reusable runtime with controllers, tools, sub-agents, session persistence, and composition tools, so you can create custom agents quickly without rebuilding the underlying machinery.