Radare2 MCP Server

An MCP server to use radare2 with AI agents such as OpenCode, Mai, VSCode, Claude, CLION, ...

Features

This implementation provides:

• 💻 Fully written in C using the native r2 APIs
• 🧩 Works from the CLI, as an r2 plugin and as an MCP server
• 🔍 Seamless binary analysis with radare2
• 🔗 Connect to any local or remote r2/iaito session via r2pipe
• 🔒 Supports readonly mode, sandbox lock and restrict tools
• 🔩 Fine grained tools configuration
• 🔁 Direct stdin/stdout communication model
• 🌐 HTTP MCP server mode with optional X-Session-ID multiplexing
• 🛠️ Optional raw access to run r2 commands or r2js scripts

Installation

Using r2pm

The simplest way to install the package is by using r2pm:

$ r2pm -Uci r2mcp

The r2mcp executable will be copied into r2pm's bindir in your home directory. However, this binary is not supposed to be executed directly from the shell; it will only work when launched by the MCP service handler of your language model of choice.

$ r2pm -r r2mcp

That's the common mcpServer JSON configuration file:

{
  "mcpServers": {
    "radare2": {
      "command": "r2pm",
      "args": ["-r", "r2mcp"]
    }
  }
}

Codex

To install the codex plugin just run the following command:

make codex-plugin-install

This will create a personal marketplace in your home and copy the files from dist/codex-plugin inside.

Using Docker

Alternatively, you can build the Docker image:

docker build -t r2mcp .

Update your MCP client configuration file (see below) to use the Docker image to use:

• "command": "docker"
• "args": ["run", "--rm", "-i", "-v", "/tmp/data:/data", "r2mcp"].

Configuration

HTTP Server Mode

Use -H [address:]port to run r2mcp as an HTTP MCP server instead of using stdin/stdout. A bare port binds only to 127.0.0.1; use 0.0.0.0:port explicitly to make the server reachable from the network:

r2mcp -H 8765

# Public binding; protect this with -a <token> or -A.
r2mcp -H 0.0.0.0:8765 -A

Per-session HTTP state is enabled with -X max[:idle_seconds] and routed by the X-Session-ID request header:

r2mcp -H 8765 -X 8:600

HTTP bearer authentication is enabled with -a <token>. Clients must send Authorization: Bearer <token> on every GET and POST request:

r2mcp -H 8765 -a my-secret-token

Use -A or -a random to generate a random token at startup. The generated token is printed once on stderr:

r2mcp -H 8765 -A

The same token setting can be supplied with R2MCP_AUTH_TOKEN; set it to random to generate one.

-X requires radare2 ABI 91 or newer because older radare2 headers do not expose the HTTP request header API. With older ABIs, -H still works as a single shared HTTP server, but -X is ignored. Bearer authentication also requires this header API.

r2 Core Plugin Mode

The r2 core plugin exposes the r2mcp command inside an existing radare2 session. Starting HTTP mode from r2 reuses the current RCore, so MCP tools operate on the file and analysis state already loaded in that session:

r2 -e r2mcp.port=8765 -c 'r2mcp start' /bin/ls

Useful subcommands:

• r2mcp start or r2mcp http starts the HTTP MCP server in the background.
• r2mcp stop, r2mcp restart, and r2mcp status manage the background server.
• r2mcp logs on|off and r2mcp logs file <path> control plugin debug logs.
• r2mcp approve on|off and r2mcp approve url <url> control supervisor approvals.
• r2mcp yolo on|off disables approvals and exposes dangerous run tools.
• r2mcp config lists the r2mcp.* eval keys.

The plugin registers eval keys for the same server options exposed by the CLI: r2mcp.port, r2mcp.log, r2mcp.logfile, r2mcp.auth, r2mcp.approve, r2mcp.svc, r2mcp.yolo, r2mcp.mini, r2mcp.permissive, r2mcp.run, r2mcp.readonly, r2mcp.ignore_analysis, r2mcp.prompts, r2mcp.prompts.dir, r2mcp.sandbox, r2mcp.sandbox.grain, r2mcp.content, r2mcp.enabled, r2mcp.disabled, r2mcp.session_tools, r2mcp.sessions, r2mcp.sessions.max, r2mcp.sessions.timeout, r2mcp.decompiler, and r2mcp.baseurl.

When loading the plugin manually with -i, set these keys after the plugin is loaded, for example with -c 'e r2mcp.port=8765', because radare2 processes early -e options before ad hoc plugin initialization.

SQL through r2xsql

Install the r2xsql core plugin for the same radare2 build that r2mcp uses. The sql tool is available without -r in local mode. Call open_file, analyze, and sql with arguments such as:

{"query":"SELECT name, size FROM funcs WHERE size > 100 ORDER BY size DESC LIMIT 10"}

The sql tool passes one statement directly to the loaded SQL core plugin, preserving SQL operators and quoted text. Results are the plugin's text output. It supports SQL writes, so it is excluded from read-only mode (-R). Use it with a local r2mcp server or r2mcp's core plugin; remote r2 HTTP client mode (-u) does not expose this tool.

An installed plugin loads automatically. To load a build explicitly, add -c 'L /absolute/path/core_r2xsql.so' to the r2mcp command. Startup commands run in order before MCP requests or -T tests, and their output is discarded to keep stdout valid for MCP. On macOS/Windows use the actual plugin filename.

Claude Desktop Integration

In the Claude Desktop app, press CMD + , to open the Developer settings. Edit the configuration file and restart the client after editing the JSON file as explained below:

1. Locate your Claude Desktop configuration file:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json

2. Add the following to your configuration file:

{
  "mcpServers": {
    "radare2": {
      "command": "r2pm",
      "args": ["-r", "r2mcp"]
    }
  }
}

VS Code Integration

To use r2mcp with GitHub Copilot Chat in Visual Studio Code by adding it to your user configuration (see other options here):

1. Open the Command Palette with CMD + Shift + P (macOS) or Ctrl + Shift + P (Windows/Linux).
2. Search for and select Copilot: Open User Configuration (typically found in ~/Library/Application Support/Code/User/mcp.json in macOS).
3. Add the following to your configuration file:

{
  "servers": {
    "radare2": {
      "type": "stdio",
      "command": "r2pm",
      "args": ["-r", "r2mcp"]
    }
  },
  "inputs": []
}

Zed Integration

You can use r2mcp with Zed as well by adding it to your configuration:

1. Open the command palette: CMD + Shift + P (macOS) or Ctrl + Shift + P (Windows/Linux).
2. Search of agent: open configuration or search of settings.
3. Add your server as such:

  "context_servers": {
    "r2-mcp-server": {
      "source": "custom",
      "command": "r2pm",
      "args": ["-r", "r2mcp"],
      "env": {}
    }
  }

Note: you will need another LLM agent, such as Claude, Gemini or else to be able to use it.

For Developers

Build from Source

Linux/macOS

To test the server locally, you can build and install it with make:

make install

This will compile the server and place the r2mcp binary in /usr/local/bin on macOS.

Windows

For Windows, just use meson and ninja like it's done in the CI:

meson b
ninja -C b