This project lets you use Ghidra in headless mode to extract rich binary analysis data (functions, pseudocode, structs, enums, etc.) into a JSON file, and expose it to LLMs like Claude via Model Context Protocol (MCP).
It turns Ghidra into an interactive reverse-engineering backend.
- Decompiles a binary using Ghidra headless mode
- Extracts:
- Function pseudocode, names, parameters, variables, strings, comments
- Data structures (structs), enums, and function definitions
- Outputs to
ghidra_context.json
- MCP server exposes tools like:
list_functions()
,get_pseudocode(name)
list_structures()
,get_structure(name)
list_enums()
,get_enum(name)
list_function_definitions()
,get_function_definition(name)
- macOS (tested)
- Python 3.10+
- Ghidra 11.3.1+
- Java 21 (Temurin preferred)
- MCP client (e.g. Claude Desktop)
mcp
CLI (install viapip install mcp
)
brew install --cask temurin@21
Then set it:
export JAVA_HOME=$(/usr/libexec/java_home -v 21)
echo 'export JAVA_HOME=$(/usr/libexec/java_home -v 21)' >> ~/.zshrc
source ~/.zshrc
Check it:
java -version
Should say: openjdk version "21.0.x"...
Download and extract Ghidra 11.3.1
cd ghidra_mcp
gcc -Wall crackme.c -o crackme
mcp install main.py
This registers the MCP server so Claude or other clients can access it.
mcp dev main.py
This enables hot reload and developer logs.
Tool | Description |
---|---|
setup_context(...) |
Run Ghidra on a binary |
list_functions() |
All functions |
get_pseudocode(name) |
Decompiled pseudocode |
list_structures() |
All structs |
get_structure(name) |
Details of a struct |
list_enums() |
All enums |
get_enum(name) |
Enum values |
list_function_definitions() |
All function prototypes |
get_function_definition() |
Return type & args |
Analyze the binary file located at <BINARY_PATH> using Ghidra installed at <GHIDRA_PATH>. First, set up the analysis context using both paths, then list all functions in the binary. Examine the main entry point function and provide a high-level overview of what the program does.
➡️ Fix: Install Java 21, not 17 or 24:
brew install --cask temurin@21
export JAVA_HOME=$(/usr/libexec/java_home -v 21)
➡️ Claude can't locate uv
by name. To fix:
- Run in your terminal:
which uv
Example output:
/Users/yourname/.cargo/bin/uv
- Open your Claude Desktop config file:
open ~/Library/Application\ Support/Claude/claude_desktop_config.json
- Update it like so:
{
"mcpServers": {
"ghidra": {
"command": "/Users/yourname/.cargo/bin/uv",
"args": [
"--directory",
"/Users/yourname/Documents/ghidra_mcp",
"run",
"main.py"
]
}
}
}
- Restart Claude Desktop. You should now see your custom MCP tools.
➡️ Fix: Java not installed or JAVA_HOME
is unset. Follow setup instructions above.
File | Purpose |
---|---|
main.py |
MCP server with tools |
export_context.py |
Ghidra script that extracts JSON |
crackme.c |
Sample C binary |
crackme |
Compiled binary to test |