Mitsu MCP server
Commit Nov 13, 2026Runs authorized penetration tests from recon to post-exploitation using 165+ connected red-team tools and persistent session intel.
Also known as ARCANGEL0/Mitsu
- Stars
- 539
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is Mitsu MCP?
Mitsu runs authorized penetration tests from recon to post-exploitation using 165+ connected red-team tools and persistent session intel.
Summary generated from the project's README
About Mitsu MCP server
Security professionals conducting authorized penetration tests can use Mitsu to automate workflows from initial reconnaissance to post-exploitation. This agentic framework acts as a red-team companion that manages an entire engagement by connecting large language models to a vast library of over 165 red-team tools. It functions as an MCP server that bridges your chosen AI model with specialized security utilities across 17 different categories, including network mapping, web application assessment, and password auditing.
Users can configure the system using various model providers such as OpenAI, Anthropic, Ollama, or DeepSeek. The framework operates in three distinct modes: Recon for passive research, Guided for operator-steered investigations, and AutoPwn for autonomous evidence-led steps. Every session maintains persistent intelligence, tracking targets, IPs, domains, and discovered vulnerabilities to build a continuous attack-surface map. Because it relies on local tool execution, the system may request approval to install missing utilities or perform actions requiring sudo authentication.
What you can do with it
- 1Automated reconnaissance and discoveryAn operator can task the agent to perform passive DNS research and public asset lookups to map an organization's footprint.
- 2Autonomous post-exploitation workflowsSecurity teams can deploy the AutoPwn mode to allow the agent to follow evidence-backed leads through various security milestones automatically.
- 3Engagement intelligence and reportingUsers can transform collected session data into a structured assessment report or an interactive map of the attack surface.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 13, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablePass
- Open source license (OSI)Pass
- Commit or release in the last yearPass
- Package name not close to a popular packageNot checked
- Remote endpoint uses HTTPSNot checked
- Destructive tools are labelledNot checked
- No hidden instructions to the modelPass
- No critical advisory in the package or its dependenciesNot checked
README
View on GitHub✦ M I T S U ✦
Your chibi red-team companion
An agentic pentest framework that can take an engagement from recon through post-exploitation on her own.
- What is Mitsu?
- At a glance
- Pentest process flow
- Demonstration
- Install
- Modes and dialogs
- Red Team Armory
- MCP integrations
- Session Intel
- Providers and search
- Using a session
- Slash commands
- Project layout
- Data storage
- Disclaimer
- License
What is Mitsu?
Mitsu is an agentic framework client for authorized penetration-testing workflows. Given an in-scope target and objective, she can carry out reconnaissance, enumerate hosts and services, research vulnerabilities and exploits, validate findings, and continue into post-exploitation. Her MCP server connects the selected model to 165+ Red Team Armory tools, external MCP servers, and custom command wrappers.
Each session keeps its own Intel, including targets, IPs, domains, services, vulnerabilities, findings, accounts, evidence, and observations. Mitsu adds to it as the engagement progresses and uses what she learns to choose what to check next. You can also turn the session Intel into an interactive attack-surface map or an assessment report, with your chibi companion along for the work.
At a glance
- Model providers: OpenAI, Anthropic, OpenRouter, Ollama, DeepSeek, Moonshot (Kimi), NVIDIA NIM, and Cloudflare Models, plus custom OpenAI- or Anthropic-compatible endpoints.
- Mitsu’s MCP server connects models to a catalog of 165+ Red Team Armory tools across 17 categories, plus custom CLI wrappers and external MCP servers.
- Persistent engagement Intel for targets, findings, services, vulnerabilities, domains, IPs, accounts, people, and working observations.
- Three operating modes: Recon, Guided, and AutoPwn.
- Engagement-scoped sessions keep conversation, Intel, artifacts, and working directory together; resume, search, rewind, or export as needed.
- Vulnerability research that combines session Intel with configured search engines and reference databases.
Mitsu Logic & Pentest Process Flow
flowchart TD
A[Launch Mitsu] --> B{Choose a session}
B -->|New| C[Create session]
B -->|Resume| D[Load history and Intel]
C --> E[Select provider and model]
D --> E
E --> F[Set target, scope, mode, and workspace]
F --> G[Describe the assessment objective]
G --> H[Model reviews conversation and session Intel]
H --> I[Choose a focused next step]
I --> J{Need a tool, a detail, or a final answer?}
J -->|Question or choice| K[Show an interactive dialog]
K --> H
J -->|Final answer| Z[Summarize evidence and remaining gaps]
J -->|Tool| L{Mode and approval check}
L -->|Guided| M[Ask operator to choose the next action]
M -->|Continue| N[Run one focused check]
M -->|Redirect or stop| H
L -->|AutoPwn| N
L -->|Recon| O[Use permitted passive or read-only checks]
L -->|Approval required| P[Show approval or sudo dialog]
P -->|Approve| N
P -->|Decline| H
N --> Q[Run capability through MCP]
O --> Q
Q --> R[Review tool output and evidence]
R --> S[Update targets, services, IPs, domains, and observations]
S --> T{Relevant vulnerability lead?}
T -->|Yes| U[Check product, version, and prerequisites]
U --> V{Evidence supports applicability?}
V -->|No or unknown| W[Record uncertainty and choose another check]
V -->|Yes| X[Validate impact within scope]
X --> Y{Confirmed?}
Y -->|No| W
Y -->|Yes| AA[Record finding, severity, evidence, and references]
T -->|No| AB{Useful in-scope work remains?}
W --> AB
AA --> AB
AB -->|Yes| I
AB -->|No| Z
Z --> AC[Save session and review results]
AC --> AD[Generate report, map attack surface, export, or resume]
AE[Operator sends a new direction] -. interrupts active turn .-> H
Mitsu runs the selected model in a tool loop: it receives the session history and Intel, requests a tool when useful, and continues from the returned evidence. Recon, Guided, and AutoPwn apply different interaction and approval rules around that loop. A new operator message can redirect an active turn; recorded Intel keeps confirmed assets and findings alongside unresolved observations.
Demonstration
Install
Mitsu requires Python 3.10 or newer. The GitHub installer uses uv to set up and manage its environment.
By pip
Install Mitsu with:
pip install mitsu
mitsu
Install from GitHub
git clone https://github.com/ARCANGEL0/Mitsu.git
cd Mitsu
./scripts/install.sh
Mitsu can identify missing tools or the local SearXNG container needed for a task and offer to install them with your approval.
Install a local checkout
From the Mitsu source directory:
uv tool install .
mitsu
To run directly from the checkout instead:
./mitsu
First launch
The first-run setup walks you through the essentials:
- Configure a model provider and its connection details.
- Choose a model.
- Configure web-search providers.
- Select the Red Team Armory tools Mitsu can use; approve installation if a selected tool is missing.
- Choose Mitsu’s companion.
After that, you’re ready to start a session.
You can return to provider and model setup from the command center at any time. In a session, use /connect to change provider settings or /model to switch models. For a local Ollama server, make sure it is running and has a model available before connecting.
Choose how Mitsu works
Switch modes with /mode. The mode controls how Mitsu interacts with targets and when she asks you to decide.
| Mode | What to expect |
|---|---|
| Recon | Passive, operator-led work such as public-source lookups, DNS research, and local read-only inspection. If the next useful step requires active interaction with a target, Mitsu opens a dialog to choose another mode or stay passive. |
| Guided | You steer the investigation. Mitsu pauses for action approval, takes a focused evidence-gathering step, records useful Intel, then asks what direction to take next. |
| AutoPwn | Mitsu chooses evidence-led next steps autonomously inside the authorized scope. She can continue through useful milestones without asking after every tool batch. Installation approval and sudo authentication still require you. |
Mitsu is designed to keep moving when a lead fails: use the evidence, drop paths that no longer fit, and choose a better-supported check. AutoPwn does not expand the target scope or authorize destructive, disruptive, persistent, or unrelated actions.
Pick the pace
Want to stay hands-off? AutoPwn lets Mitsu follow useful, evidence-backed paths within the scope you set, with minimal interruptions. Prefer to direct the engagement? In Guided, Mitsu reports what she found, lays out the next useful directions, and waits for you to choose where she goes. You’re the orchestrator; she brings the chibi rage to the target you picked.
Either way, Mitsu keeps the objective and recorded evidence in view, drops leads that stop making sense, and pauses when an action needs your approval or credentials.
The Red Team Armory
Mitsu comes equipped with an Armory catalog of 165+ red-team tools across 17 categories, exposed through her MCP server with descriptions and inputs the model can use. You can expand her kit whenever you need by registering custom command wrappers or connecting additional MCP servers. When a catalogued utility is missing, Mitsu can offer to install it and wait for your approval.
| Category | Catalog | Typical work |
|---|---|---|
| OSINT / Recon | 22 | Public assets, domains, DNS, and organization research |
| Network Mapping | 21 | Host and service discovery within scope |
| Web Application Assessment | 21 | Web app and CMS checks |
| API Security Assessment | 4 | REST, GraphQL, and token-based API checks |
| Database Assessment | 2 | Database exposure and security review |
| Vulnerability Assessment | 8 | Known exposures and security posture |
| Password Auditing | 14 | Approved hashes and authentication controls |
| Exploitation Frameworks | 3 | Authorized vulnerability validation |
| Active Directory | 11 | Windows domain and identity-path assessment |
| Post-Exploitation & Pivoting | 9 | Lab access, privilege boundaries, and routes |
| Traffic Analysis & Spoofing | 6 | Authorized traffic inspection |
| Wireless Security | 6 | Approved wireless assessments |
| Hardware & Radio | 5 | Authorized Bluetooth, SDR, and radio work |
| Social Engineering Simulations | 2 | Consent-based exercises |
| Cloud & Container Assessment | 7 | Cloud, container, and Kubernetes configuration |
| Mobile Application Assessment | 4 | Android and mobile test environments |
| Forensics / Reverse Engineering | 20 | Supplied binaries, firmware, and evidence |
OSINT / Recon (22)
OWASP Amass, Subfinder, assetfinder, theHarvester, Recon-ng, SpiderFoot, Sherlock, Maigret, DNSRecon, DNSenum, Fierce, DMitry, WHOIS, gau, waybackurls, Katana, CloudEnum, Shodan CLI, Censys CLI, ASNMap, URLFinder, tlsx
Network Mapping (21)
Nmap, Masscan, ZMap, RustScan, arp-scan, Netdiscover, fping, hping3, Naabu, dnsx, httpx (ProjectDiscovery), Uncover, AutoRecon, NBTScan, enum4linux, SMBMap, rpcclient, Net-SNMP Walk, Traceroute, Netcat, socat
Web Application Assessment (21)
Burp Suite Community, OWASP ZAP, WPScan, sqlmap, Nikto, Gobuster, ParamSpider, hakrawler, GoSpider, Jaeles, x8, ffuf, feroxbuster, dirsearch, DIRB, WhatWeb, WAFW00F, Arjun, Dalfox, Commix, XSStrike
API Security Assessment (4)
JWT Tool, GraphQLmap, grpcurl, Schemathesis
Database Assessment (2)
NoSQLMap, sqlninja
Vulnerability Assessment (8)
Nuclei, testssl.sh, sslscan, SSLyze, Trivy, Vulnx, Checkov, Grype
Password Auditing (14)
John the Ripper, Hashcat, THC Hydra, Medusa, Ncrack, HashID, Crunch, Patator, Hash Identifier, Ophcrack, Stegseek, Crowbar, fcrackzip, CeWL
Exploitation Frameworks (3)
Metasploit Framework, Exploit-DB / SearchSploit, RouterSploit
Active Directory (11)
NetExec, Impacket, BloodHound.py, Certipy, enum4linux-ng, Kerbrute, ldapdomaindump, ADIDNSDump, bloodyAD, Coercer, pypykatz
Post-Exploitation & Pivoting (9)
Evil-WinRM, Responder, Ligolo-ng, Chisel, ProxyChains-NG, sshuttle, pspy, LinPEAS, WinPEAS
Traffic Analysis & Spoofing (6)
tcpdump, Wireshark, TShark, dsniff, Ettercap, mitmproxy
Wireless Security (6)
Aircrack-ng, Bettercap, Kismet, Wifite, hcxdumptool, hcxtools
Hardware & Radio (5)
BlueZ, HackRF Tools, RTL-SDR Tools, Gqrx SDR, Ubertooth Tools
Social Engineering Simulations (2)
Social-Engineer Toolkit, Gophish
Cloud & Container Assessment (7)
Prowler, ScoutSuite, kube-hunter, Pacu, ROADrecon, CloudFox, kube-bench
Mobile Application Assessment (4)
Apktool, JADX, Frida Tools, Objection
Forensics / Reverse Engineering (20)
Ghidra, Radare2, Volatility 3, Binwalk, Foremost, ExifTool, YARA, GNU Debugger, GDB Multiarch, Rizin, checksec, ROPgadget, Ropper, pwntools, Steghide, zsteg, TestDisk / PhotoRec, The Sleuth Kit, bulk_extractor, Scalpel
Browse the catalog and manage installations from the Red Team Armory in the command center. Tool availability depends on what is installed and supported on the host; the session status shows how many selected tools are ready to use.
MCP integrations
Mitsu includes an MCP server that connects the agent to built-in session capabilities and external tools. Configure it with /mcp or from Configure MCP Server in the command center.
The configuration supports:
- Red Team Armory: enable or disable the catalog and select individual tools or categories.
- External MCP servers: connect native stdio or HTTP servers.
- Custom CLI tools: describe a local executable, its purpose, arguments, and examples so it can be exposed as a named MCP tool.
A minimal configuration looks like this:
{
"servers": [
{
"name": "local-tools",
"command": "python",
"args": ["-m", "example_mcp_server"],
"env": {}
}
],
"custom_tools": [],
"armory": {
"enabled": true,
"tools": ["network-mapping", "vulnerability"]
}
}
Use absolute executable paths when a server is not available on Mitsu’s PATH. Keep credentials out of command arguments and transcripts; use the server’s supported environment configuration for secrets.
Wrap a CLI utility as an MCP tool
Custom tool definitions give Mitsu a named executable, documentation, and typed inputs instead of a free-form shell string.
{
"id": "staging_audit",
"name": "Staging audit",
"command": "/usr/local/bin/staging-audit",
"description": "Run the approved read-only audit against a staging host.",
"usage": "staging-audit --target HOST",
"examples": ["Audit the authorized staging host"],
"args": [],
"parameters": [
{
"name": "target",
"type": "string",
"description": "Authorized staging hostname",
"flag": "--target",
"required": true
}
],
"env": {},
"timeout_seconds": 120,
"read_only": true
}
The full configuration is saved in ~/.config/mitsu/mcp_servers.json by default. Mitsu validates tool IDs, parameter types, required fields, timeouts, and selected Armory entries before reconnecting.
Session Intel
Each session carries its own compact Intel dataset. Mitsu saves it inside that session’s JSON file, reloads it with the conversation, and supplies it as context while she works. It is ordinary session data, not a separate SQLite database.
The TUI keeps Intel in tabs on the right side of the session. Click a tab to view that category’s records:
┌─ Session: target assessment ─────────────────────────────┐
│ Conversation · scope · working directory │
│ │
│ Session Intel → loaded as context for Mitsu’s next turn │
│ │
│ Right-side Intel tabs │
│ [Targets] [Findings] [Services] [Exploits / Vulns] │
│ [Domains] [IPs] [Passwords / Accounts] [Manhunt] │
│ [Observations] │
│ │
│ Select a tab to browse the session’s records │
└──────────────────────────────────────────────────────────┘
By default, a session is stored at ~/.local/share/mitsu/sessions/<session-id>.json; setting XDG_DATA_HOME changes the base directory. The same file contains the conversation and session Intel, so each assessment keeps its own target context when you resume it.
| Intel section | What belongs there |
|---|---|
| Targets | Authorized hosts, domains, URLs, and other assessment scope |
| Findings | Useful confirmed information and vulnerability leads, with evidence and status |
| Services | Confirmed ports, protocols, products, versions, and banners |
| Exploits / Vulnerabilities | CVEs, affected-version notes, references, and validation state |
| Domains | Domains, hostnames, nameservers, and related naming information |
| IPs | Addresses discovered in the authorized environment |
| Passwords / Accounts | Users, groups, credentials, tokens, hashes, and lab flags |
| Manhunt | Person-scoped OSINT such as names, emails, phone numbers, and social profiles |
| Observations | Draft and validated notes about what worked, what failed, and what to try next |
Records can include their source, evidence, confidence, verification state, timestamps, and relationships to other assets. A scan output is not automatically proof that every result is exploitable: vulnerability leads stay leads until the affected version and target behavior support them.
Mitsu can read and update Intel through MCP during a session. Older session data is migrated into the current structured format when loaded.
Models and providers
Mitsu connects to the OpenAI and Anthropic APIs directly. Other built-in connections include OpenRouter, NVIDIA NIM, Ollama, DeepSeek, and Moonshot (Kimi) through OpenAI-compatible APIs, plus Cloudflare Models through a Cloudflare account endpoint. You can also configure a custom endpoint that follows either the OpenAI or Anthropic API format.
Image input is supported for models that accept vision. Paste an image into chat or attach one from the session workspace; Mitsu checks the selected model’s image capability before sending it. Supported formats include PNG, JPEG, WebP, and GIF. Image support depends on the provider and specific model, so choose a vision-capable model when working from screenshots or other images.
API keys and provider profiles are stored in your local Mitsu configuration. When using a hosted provider, prompts, tool results, Intel context, and any attached images needed for the request are sent to that provider.
Web search
Configure search sources with /config. Mitsu can query multiple configured engines and combine results. Available options include direct search providers, API-backed engines, and SearXNG. Some engines require an API key; SearXNG requires Docker for Mitsu’s managed local setup.
| Connection type | Search engines |
|---|---|
| Direct search, no key | DuckDuckGo, Google, Bing, Brave, Mojeek, Ecosia, Yahoo, Startpage |
| API key or account | Exa, Tavily, Ollama Web Search, Brave Search API, Serper, SerpApi, Firecrawl, Google Programmable Search |
| Self-hosted metasearch | SearXNG |
Web search is useful for current documentation, public records, and vulnerability references. Search results are leads: confirm affected versions and target-side behavior before treating a CVE as applicable.
Vulnerability search
Use /vulnsearch <query> to look for vulnerability and exploit references. With no query, Mitsu can search using product and version clues already recorded in Intel. Sources can include:
- Saved session Intel.
- Configured web search engines.
- Vulnx, when installed.
- Exploit-DB references from the local SearchSploit database, when available.
- Direct NVD/NIST database search is not integrated yet; it is on the future TODO list.
The results are presented as findings and references. A match is not proof that the target is vulnerable, and the search does not run an exploit.
Using a session
Start a new session from the command center and describe the task in plain language. Include the authorized scope, any known addresses or domains, and the objective. You can attach useful context or point Mitsu to files in the session workspace.
Example:
I’m assessing the authorized host at 192.0.2.25. Start with the exposed services, record confirmed ports and versions in Intel, then suggest the most relevant next check.
Set the working directory before asking Mitsu to create or inspect engagement files:
/workdir /path/to/engagement
The working directory is stored with that session and becomes its primary workspace for file and shell operations. Mitsu starts in the project’s current directory when a session has no custom workdir configured.
During a run, the chat shows tool activity and expandable command output. You can stop a response, copy messages or code blocks, rewind to an earlier conversation point, and search the chat. Image attachments can be added through paste or file references.
Mitsu MCP questions
What is Mitsu?
Mitsu is an agentic pentest framework that can take an authorized engagement from reconnaissance and attack-surface mapping through vulnerability discovery, exploitation, and post-exploitation.
How many tools does Mitsu's MCP server connect?
Her MCP server connects a model to a catalog of more than 165 red-team tools across 17 categories, plus custom CLI wrappers and external MCP servers.
Which operating modes does Mitsu support?
Mitsu offers Recon, Guided, and AutoPwn modes. Recon is passive and operator-led, Guided pauses for approval on each focused step, and AutoPwn picks evidence-led next steps autonomously within scope.
Which model providers can Mitsu use?
Mitsu supports OpenAI, Anthropic, OpenRouter, Ollama, DeepSeek, Moonshot (Kimi), NVIDIA NIM, and Cloudflare Models, plus custom OpenAI- or Anthropic-compatible endpoints.
Does AutoPwn expand the target scope?
No. AutoPwn follows evidence-backed paths inside the authorized scope and does not authorize destructive, disruptive, persistent, or unrelated actions; installation approval and sudo authentication still require the operator.