Mythic MCP
# AI-Powered Red Teaming with Claude, C2 & MCP
A project exploring the integration of *Claude, Model Context Protocol (MCP), and Command & Control (C2) Mythic* to build an AI-powered copilot for Red Team operations.
The project aims to enable Claude to interact with C2 capabilities through MCP, understand the context and results of an ongoing operation, and provide *context-aware recommendations, attack paths, and next-step actions* based on the information gathered during the engagement.
## Integration Diagram
## Requirements
1. uv
2. python3
3. Claude Desktop
4. Mythic C2 (Use docker to run the service)
5. A Windows victim machine in VMware
## Preliminary steps
1. Install and run the Mythic C2 server. You need to use the official repository and run it in Docker. Repository link: [Mythic Repository](https://github.com/its-a-feature/Mythic)
2. Install the agents (Apollo, Apfell, Poseidon) and the profiles (HTTP, WebSocket).
3. Use some of the payloads to generate a payload to send to the victim machine.
4. On the victim machine, execute the payload to obtain a session on the Mythic C2 server.
## Usage with Claude Desktop
To deploy an MCP server with Claude Desktop, we need to use the Developer Options. You will need to edit the *claude_desktop_config.json* file to configure the MCP server. The configuration format is as follows:
```
{
"mcpServers": {
"mythic_mcp": {
"command": "/snap/bin/uv",
"args": [
"--directory",
"/path/to/mythic_mcp/",
"run",
"main.py",
"mythic_admin",
"mythic_admin_password",
"localhost",
"7443"
]
}
}
}
```
To find the `mythic_admin` and `mythic_admin_password` credentials after deploying the Mythic server, you can use the following commands:
```
sudo ./mythic-cli config get admin_user
sudo ./mythic-cli config get admin_password
```
## How to Use
Once you have an active session in C2 and have uploaded the configuration file to Claude, you can use `any prompt`. However, we recommend running a few verification prompts first to ensure that the connection between C2, Claude, and the MCP server is working correctly.
#### Verify the available sessions
```
How many sessions are currently available?
```
#### Verify the available commands
```
Which commands are available?
```
TDQS
Scored across 12 tools
Most tools target distinct actions (file read vs. upload, shell vs. PowerShell, mimikatz vs. Rubeus), but ad_recon and enum_domain_users overlap in purpose, as enum_domain_users is a specific subset of AD reconnaissance. This is a minor ambiguity rather than a major one.
The majority follow a verb_noun pattern (read_file, run_shell_command, execute_powershell, upload_file), but ad_recon and privilege_escalation_peas deviate by using nouns or abbreviations without a clear verb. The inconsistency is noticeable but not chaotic.
With 12 tools, the set is well-scoped for a post-exploitation/Mythic C2 server. Each tool serves a clear function in the red-team workflow, and the count fits comfortably within the ideal range without feeling bloated or sparse.
The set covers core post-exploitation activities: execution, file operations, credential theft, privilege escalation, and AD reconnaissance. Missing features like process listing or lateral movement are notable but not critical for the primary purpose, and agents can often work around these gaps.