pyghidra-mcp is a command-line Model Context Protocol (MCP) server that brings the full analytical power of Ghidra, a robust software reverse engineering (SRE) suite, into the world of intelligent agents and LLM-based tooling.
It bridges Ghidra’s ProgramAPI and FlatProgramAPI to Python using pyghidra and jpype, then exposes that functionality via the Model Context Protocol.
MCP is a unified interface that allows language models, development tools (like VS Code), and autonomous agents to access structured context, invoke tooling, and collaborate intelligently. Think of MCP as the bridge between powerful analysis tools and the LLM ecosystem.
With pyghidra-mcp, Ghidra becomes an intelligent backend—ready to respond to context-rich queries, automate deep reverse engineering tasks, and integrate into AI-assisted workflows.
pyghidra-mcp now supports two operating modes:
headlessmode for CLI-driven analysis and automation--guimode, which launches Ghidra throughpyghidra-mcpand shares live program state with the running GUI
Note
This beta project is under active development. We would love your feedback, bug reports, feature requests, and code.
Yes, the original ghidra-mcp is fantastic. But pyghidra-mcp takes a different approach:
- 🐍 Headless-first, GUI-capable – Run entirely via CLI for streamlined automation, or launch Ghidra with
--guiwhen you want live GUI navigation and edits. - 🔁 Designed for automation – Ideal for integrating with LLMs, CI pipelines, and tooling that needs repeatable behavior.
- ✅ CI/CD friendly – Built with robust unit and integration tests for both client and server sessions.
- 🚀 Quick startup – Asynchronous startup allows the server to start handling requests while binaries are still being analyzed in the background. Supports fast command-line launching with minimal setup.
- 📦 Project-wide analysis – Enables concurrent reverse engineering of all binaries in a Ghidra project
- 🤖 Agent-ready – Built for intelligent agent-driven workflows and large-scale reverse engineering automation.
- 🔍 Semantic code search – Uses vector embeddings (via ChromaDB) to enable fast, fuzzy lookup across decompiled functions, comments, and symbols—perfect for pseudo-C exploration and agent-driven triage.
This project provides a Python-first experience optimized for local development, headless environments, and testable workflows.
flowchart LR
subgraph Clients["Clients"]
Agent["MCP host / agent"]
Cli["pyghidra-mcp-cli"]
User["Ghidra user"]
end
subgraph Process["pyghidra-mcp process"]
Transport["streamable HTTP (recommended)<br/>stdio (local fallback)"]
Tools["MCP tools"]
Context["PyGhidra context"]
end
Project["Ghidra project<br/>.gpr / .rep"]
Artifacts["MCP artifacts<br/>ChromaDB + GZF cache"]
Gui["Ghidra GUI / CodeBrowser<br/>only with --gui"]
Agent -->|"HTTP (recommended)"| Transport
Cli -->|"HTTP only"| Transport
Transport --> Tools
Tools --> Context
Context --> Project
Context --> Artifacts
Context -.-> Gui
User -.-> Gui
Gui -.-> Project
flowchart TD
Start["What do you need?"]
Start --> Headless["Agent or automation"]
Start --> GuiNeed["Live Ghidra GUI control"]
Start --> Terminal["Interactive terminal client"]
Headless --> Http["pyghidra-mcp<br/>--transport streamable-http"]
GuiNeed --> GuiMode["pyghidra-mcp --gui<br/>--transport streamable-http<br/>--project-path project.gpr"]
Terminal --> HttpServer["Start pyghidra-mcp<br/>--transport streamable-http"]
HttpServer --> CliMode["Run pyghidra-mcp-cli commands"]
- Headless MCP: use
streamable-http. It keeps Ghidra and its project alive as one long-running server and works with both MCP hosts and the CLI client. - GUI mode:
pyghidra-mcplaunches Ghidra, opens the project, and exposes extra tools that steer the CodeBrowser in the same JVM. - CLI client:
pyghidra-mcp-cliis an HTTP client. Start astreamable-httpserver first, then issue terminal commands against that running server. - stdio: retain this only for MCP hosts that cannot connect to an HTTP endpoint.
Detailed architecture and tool surface
flowchart TD
subgraph Clients
Agent["LLM / MCP host"]
Cli["pyghidra-mcp-cli"]
Automation["scripts and CI"]
end
subgraph Transports
Stdio["stdio"]
Http["streamable-http"]
Sse["sse legacy"]
end
subgraph Server["pyghidra-mcp server"]
McpServer["MCPServer tool server"]
Context["PyGhidra context"]
Indexing["background analysis and Chroma indexing"]
subgraph Tools["MCP tools"]
Analysis["decompile, xrefs, bytes, callgraph"]
Search["symbols, strings, code"]
ProjectOps["import, delete, metadata, list binaries"]
Edits["rename function, rename variable, set type, set prototype, set comment"]
GuiOnly["GUI only: open program, goto, list open programs, set current program"]
end
end
subgraph GhidraRuntime["Ghidra runtime"]
PyGhidra["pyghidra"]
Jpype["JPype shared JVM"]
Project["Ghidra project"]
Programs["program databases"]
CodeBrowser["Ghidra GUI / CodeBrowser"]
end
Agent --> Http
Automation --> Http
Automation --> Sse
Cli --> Http
Stdio --> McpServer
Http --> McpServer
Sse --> McpServer
McpServer --> Context
Context --> PyGhidra
PyGhidra --> Jpype
Jpype --> Project
Project --> Programs
Context --> Indexing
Indexing --> Search
McpServer --> Tools
Tools --> Context
GuiOnly -.-> CodeBrowser
Context -.-> CodeBrowser
- PyGhidra-MCP - Ghidra Model Context Protocol Server
Start a persistent Streamable HTTP server using the Python package and uv:
uvx pyghidra-mcp \
--transport streamable-http \
--project-path /absolute/path/to/ghidra-projects \
/absolute/path/to/binaryThe server is then available at http://127.0.0.1:8000/mcp. Leave it running and configure your MCP client to use that URL; the Claude Code and Codex examples below show the expected configuration.
To launch and control a live Ghidra GUI from MCP, use --gui with streamable-http:
uvx pyghidra-mcp \
--gui \
--transport streamable-http \
--host 127.0.0.1 \
--port 8000 \
--project-path /absolute/path/to/ghidra-projects \
--project-name my_projectImportant
--gui launches Ghidra through pyghidra-mcp. It does not attach to an already-running external Ghidra instance.
Or, run as a Docker container:
docker run --rm -p 8000:8000 ghcr.io/clearbluejar/pyghidra-mcppyghidra-mcp keeps the MCP surface intentionally narrow so agent clients spend fewer tokens on tool discovery and argument selection.
- Short tool descriptions: MCP tool docstrings are kept compact so MCPServer tool schemas stay small and cheap to send to models.
- Context discipline: tools return focused structured data instead of dumping whole-program context by default. Decompilation, symbol search, and cross-reference results are shaped to support iterative analysis rather than one large response.
- GUI tools only when relevant: GUI-only controls such as
open_program_in_gui,list_open_programs,set_current_program, andgotoare only exposed when the server is started with--gui. - CLI is optional: if MCP is not your preferred interface,
pyghidra-mcp-cliprovides a direct command-line client over HTTP with grouped commands for common edit and analysis workflows.
This keeps the default server usable for LLM agents, IDE integrations, and automation without exposing unnecessary tool surface or GUI-only controls in headless sessions.
For a more interactive command-line experience, you can use the separate pyghidra-mcp-cli package, which provides a user-friendly interface for interacting with a running pyghidra-mcp server.
Install the CLI client using uv (recommended):
uvx pyghidra-mcp-cliOr install with pip:
pip install pyghidra-mcp-cli- Start the server (in one terminal):
pyghidra-mcp --transport streamable-http /bin/ls- Use the CLI (in another terminal):
# List available binaries
pyghidra-mcp-cli list binaries
# Decompile a function
pyghidra-mcp-cli decompile --binary ls main
# Decompile with callees, referenced strings, and cross-references
pyghidra-mcp-cli decompile --binary ls main --callees --strings --xrefs
# Search for symbols (supports regex patterns)
pyghidra-mcp-cli search symbols --binary ls printf -l 10Note
The CLI connects to pyghidra-mcp via HTTP to avoid the 10-60 second startup overhead of spawning a new Ghidra process for each command. See the CLI README for complete documentation.
The server and optional CLI support mcp>=1.26.0,<3. Install the same packages
with either SDK major; no SDK-specific edition or runtime flag is required.
The server selects SDK 1's FastMCP or SDK 2's MCPServer from the installed
dependency, and the CLI accepts both HTTP transport return formats.
Fresh installations normally resolve to the latest compatible SDK. To keep an existing SDK 1 environment, install with an explicit constraint:
pip install pyghidra-mcp 'mcp>=1.26.0,<2'Python SDK versions are separate from the MCP protocol versions negotiated by clients and servers.
You can create new projects in several ways, depending on your workflow:
pyghidra-mcp creates a self-contained project structure where each project has its own Ghidra project and pyghidra-mcp artifacts. This ensures complete isolation and easy project management.
# Create a new project with default settings
pyghidra-mcp
# Creates:
$ tree pyghidra_mcp_projects/
pyghidra_mcp_projects/
├── my_project.gpr
├── my_project-pyghidra-mcp
│ ├── chromadb
│ └── gzfs
└── my_project.rep# Create project with custom name and location
pyghidra-mcp --project-path ~/analysis/malware_study --project-name malware_analysis
$ tree ~/analysis/
/home/vscode/analysis/
└── malware_study
├── malware_analysis.gpr
├── malware_analysis-pyghidra-mcp
│ ├── chromadb
│ └── gzfs
└── malware_analysis.rep# Create separate projects for different analysis focuses
mkdir ~/reverse_engineering_workspace
# Project for suspicious binaries
pyghidra-mcp --project-path ~/reverse_engineering_workspace/suspicious_binaries --project-name suspicious_analysis
# Project for packed malware
pyghidra-mcp --project-path ~/reverse_engineering_workspace/packed_malware --project-name packed_analysisIf you have existing Ghidra projects (.gpr files), you can open them directly with pyghidra-mcp:
# Open existing Ghidra project (project name derived from filename)
pyghidra-mcp --project-path ~/existing/ghidra/my_research.gpr
# Result: ~/existing/ghidra/my_research-pyghidra-mcp/
# └── chromadb/, gzfs/ (pyghidra-mcp additions)Use GUI mode when you want MCP actions to operate against the same live program objects that Ghidra is displaying.
--guirequires--transport streamable-http(or--transport httpas an alias)--project-pathcan be a project directory plus--project-name, or an existing.gprfile. Missing projects are created automatically.- Ghidra is launched by
pyghidra-mcp, which keeps GUI and MCP transactions in the same JVM - GUI-only tools are only exposed when running with
--gui
Example:
pyghidra-mcp \
--gui \
--transport streamable-http \
--project-path /absolute/path/to/my_research.gprGUI mode is the right choice when you want to:
- open or switch programs in CodeBrowser
- navigate the listing to a function or address
- rename functions or add comments and immediately see those changes in Ghidra
pyghidra-mcp does not require --wait-for-analysis by default. The server can start while analysis and MCP-side indexing continue in the background.
This matters for large projects:
- starting a project with many binaries does not need to block server startup
--wait-for-analysisis available when you want a fully analyzed project before serving requests- for large existing projects, expect analysis and indexing readiness to vary by binary
Current limitation:
- Ghidra analysis state and MCP indexing state are separate
- a binary can be fully analyzed in Ghidra while
search_stringsor semanticsearch_codeare still waiting on MCP-side indexing - this is more noticeable when opening larger existing projects
In practice:
- decompilation, navigation, renaming, and comments can still work for a binary while indexing-heavy search features are catching up
- if startup latency matters more than immediate search readiness, keep the default
--no-wait-for-analysis - if immediate readiness matters more than startup time, use
--wait-for-analysis
This project uses a Makefile to streamline development and testing. ruff is used for linting and formatting, and pre-commit hooks are used to ensure code quality.
-
Install
uv: If you don't haveuvinstalled, you can install it using pip:pip install uv
Or, follow the official
uvinstallation guide: https://docs.astral.sh/uv/install/ -
Create a virtual environment and install dependencies:
make dev-setup source ./.venv/bin/activate -
Set Ghidra Environment Variable: Download and install Ghidra, then set the
GHIDRA_INSTALL_DIRenvironment variable to your Ghidra installation directory.# For Linux / Mac export GHIDRA_INSTALL_DIR="/path/to/ghidra/" # For Windows PowerShell [System.Environment]::SetEnvironmentVariable('GHIDRA_INSTALL_DIR','C:\path\to\ghidra')
The Makefile provides several targets for testing and code quality:
make run: Run the MCP server.make test: Run the full test suite (unit and integration).make test-unit: Run unit tests.make test-integration: Run integration tests.make test-integration-fast: Run the lightweight integration smoke test used by pre-commit.make test-integration-gui: Run GUI integration tests. Requires a working Ghidra install and GUI support.make lint: Check code style withruff.make format: Format code withruff.make typecheck: Run lightweight static checks withruff.make check: Run all quality checks.make dev: Run the development workflow (format and check).make build: Build distribution packages.make clean: Clean build artifacts and cache.
Recommended split:
- pre-commit:
ruff,pyright, unit tests, and one lightweight integration smoke test - GitHub Actions: full Linux headless integration coverage, Linux GUI under
Xvfb, CLI coverage, and current macOS smoke tests - scheduled CI: older macOS / Ghidra compatibility coverage
- local/manual: heavier environment-specific GUI debugging and release sanity checks
Enable LLMs to perform actions, make deterministic computations, and interact with external services.
decompile_function and list_xrefs accept a single target or a list of targets, reducing round-trips when analyzing call chains or multiple symbols at once.
Per-item errors are returned inline (other targets still succeed):
[
{"name": "main", "code": "void main() { ... }", "callees": ["init_hardware"], "xrefs": [...]},
{"name": "0xdeadbeef", "code": "", "error": "Function or symbol '0xdeadbeef' not found."}
]-
search_code(binary_name: str, query: str, limit: int = 5, offset: int = 0, search_mode: str = "semantic", include_full_code: bool = True, preview_length: int = 500, similarity_threshold: float = 0.0): Search decompiled pseudo-C using semantic vector search or literal matching. -
list_xrefs(binary_name: str, name_or_address: str | list[str]): List cross-references to function(s), symbol(s), or address(es). Accepts a single target or a list for batch lookup. -
gen_callgraph(binary_name: str, function_name: str, direction: str = "calling", display_type: str = "flow", condense_threshold: int = 50, top_layers: int = 3, bottom_layers: int = 3, max_run_time: int = 120): Generates a MermaidJS call graph for a specified function. Supports both "calling" (functions called by the target) and "called" (functions that call the target) directions with multiple visualization types. -
decompile_function(binary_name: str, name_or_address: str | list[str], include_callees: bool = False, include_strings: bool = False, include_xrefs: bool = False, timeout_sec: int = 30): Decompile function(s) by name or address. Accepts a single target or a list for batch decompilation. Rich response flags attach callees, strings, and/or xrefs to each result.timeout_secapplies per target and bounds each decompilation attempt independently. -
list_exports(binary_name: str, query: str = ".*", offset: int = 0, limit: int = 25): Lists all exported functions and symbols from a specified binary (regex supported for query). -
list_imports(binary_name: str, query: str = ".*", offset: int = 0, limit: int = 25): Lists all imported functions and symbols for a specified binary (regex supported for query). -
read_bytes(binary_name: str, address: str, size: int = 32): Reads raw bytes from memory at a specified address. Hex addresses may include or omit the0xprefix. -
search_strings(binary_name: str, query: str, limit: int = 100): Searches strings within a binary. -
search_symbols_by_name(binary_name: str, query: str, functions_only: bool = False, offset: int = 0, limit: int = 25): Search for symbols within a binary by name. Supports regex patterns (e.g.^main$,func.*one) with case-insensitive matching, or plain substring queries. Setfunctions_only=Trueto exclude labels, variables, and other non-function symbols.
-
import_binary(binary_path: str): Imports a binary from a designated path into the current Ghidra project. If the path is a directory, it will recursively scan and import all supported binary files, preserving the directory structure within the Ghidra project. -
list_project_binaries(): Lists binaries in the current Ghidra project. In GUI mode this includes project binaries that exist on disk even if they are not currently open in CodeBrowser. -
list_project_binary_metadata(binary_name: str): Retrieves detailed metadata for a specific binary, including architecture, compiler, executable format, analysis metrics, and file hashes. -
delete_project_binary(binary_name: str): Deletes a binary (program) from the Ghidra project.
-
rename_function(binary_name: str, name_or_address: str, new_name: str): Rename a function by name or address. In GUI mode this runs as a live Ghidra transaction and updates the open program. -
rename_variable(binary_name: str, function_name_or_address: str, variable_name: str, new_name: str): Rename a function parameter or local variable by exact name within a specific function. If the name is missing or ambiguous within that function, the tool returns an error instead of guessing. In GUI mode this runs as a live Ghidra transaction and updates the open program. -
set_variable_type(binary_name: str, function_name_or_address: str, variable_name: str, type_name: str): Set the data type for a function parameter or local variable by exact name within a specific function. If the name is missing or ambiguous within that function, the tool returns an error instead of guessing.type_nameis parsed using Ghidra's datatype parser against the program datatype manager. -
set_function_prototype(binary_name: str, function_name_or_address: str, prototype: str): Set a function prototype from a full signature string. The tool always runs the prototype through Ghidra's native signature parser and returns the underlying parser or apply error if the prototype is invalid. -
set_comment(binary_name: str, target: str, comment: str, comment_type: str): Set a function/decompiler comment or listing comment. Listing comment targets can be addresses, symbols, or functions. Supportedcomment_typevalues aredecompiler,plate,pre,eol,post, andrepeatable.
These tools are only available when pyghidra-mcp is started with --gui and control what the GUI is showing rather than mutating project data directly:
list_open_programs(): List programs currently open in the Ghidra GUI.open_program_in_gui(binary_name: str, new_window: bool = True): Open a project binary in CodeBrowser. By default this opens a new CodeBrowser window. Setnew_window=falseto reuse a visible CodeBrowser when possible.set_current_program(binary_name: str): Make an open program the active/current program in the primary GUI tool context.goto(binary_name: str, target: str, target_type: str): Navigate the Ghidra GUI to an address or function.target_typemust beaddressorfunction.
This Python package is published to PyPI as pyghidra-mcp and can be installed and run with pip, pipx, uv, poetry, or any Python package manager.
$ uvx pyghidra-mcp --help
Usage: pyghidra-mcp [OPTIONS] [INPUT_PATHS]...
PyGhidra Command-Line MCP server
Options:
-v, --version Show version and exit.
-t, --transport [stdio|streamable-http|sse|http]
Transport protocol. SSE is deprecated;
use streamable-http instead. [default: stdio]
-p, --port INTEGER Port for HTTP-based transports. [default: 8000]
-o, --host TEXT Host for HTTP-based transports. [default: 127.0.0.1]
--project-path PATH Directory for a pyghidra-mcp project or an
existing Ghidra .gpr file. [default: pyghidra_mcp_projects]
--project-name TEXT Ghidra project name. Ignored for .gpr paths.
[default: my_project]
--threaded / --no-threaded Allow threaded analysis. [default: threaded]
--max-workers INTEGER Number of analysis workers; 0 means CPU count.
[default: 0]
--wait-for-analysis / --no-wait-for-analysis
Wait for initial analysis before starting.
[default: no-wait-for-analysis]
--gui / --no-gui Launch Ghidra GUI in-process and serve MCP
against GUI-open programs. Cannot attach to
an already-running external Ghidra process.
[default: no-gui]
--list-project-binaries List ingested project binaries and exit.
--delete-project-binary TEXT Delete a project binary by name and exit.
--force-analysis / --no-force-analysis
Force a new binary analysis each run.
[default: no-force-analysis]
--verbose-analysis / --no-verbose-analysis
Verbose logging for analysis. [default: no-verbose-analysis]
--no-symbols / --with-symbols Turn off symbols for analysis. [default: with-symbols]
--sym-file-path PATH Single PDB symbol file for one binary.
-s, --symbols-path PATH Local symbols directory.
--gdt PATH Path to GDT files. May be specified multiple times.
--program-options PATH JSON file with Ghidra program options.
--gzfs-path PATH Location to store GZFs of analyzed binaries.
-h, --help Show this message and exit.
When using the Docker container, you can map a local directory containing your binaries into the container's workspace. This allows pyghidra-mcp to analyze your files.
# Create and populate the new directory
mkdir -p ./binaries
cp /path/to/your/binaries/* ./binaries/
# Run the Docker container with volume mapping
docker run -i --rm \
-v "$(pwd)/binaries:/binaries" \
ghcr.io/clearbluejar/pyghidra-mcp \
/binaries/*Streamable HTTP is the recommended transport. It keeps one Ghidra process and project available to MCP hosts and pyghidra-mcp-cli, avoiding the startup cost of a separate process per client. It sends JSON-RPC requests over HTTP; see the spec for details.
By default, the server listens on http://127.0.0.1:8000/mcp for client connections. Use --host / --port or the MCP_HOST / MCP_PORT environment variables to change the bind address. The server must be running for clients to connect to it.
pyghidra-mcp -t streamable-httpThe Python package defaults to stdio, so always include --transport streamable-http when starting the server.
GUI mode uses this transport:
pyghidra-mcp \
--gui \
--transport streamable-http \
--project-path /absolute/path/to/my_project.gprdocker run -p 8000:8000 ghcr.io/clearbluejar/pyghidra-mcp
After starting the HTTP server, add it to Claude Code:
claude mcp add --transport http pyghidra-mcp http://127.0.0.1:8000/mcpFor a project-shared configuration, add this to .mcp.json in the project root:
{
"mcpServers": {
"pyghidra-mcp": {
"type": "http",
"url": "http://127.0.0.1:8000/mcp"
}
}
}type: "http" is required: an entry with only url is treated as a stdio server by Claude Code. Use claude mcp list or /mcp to verify the connection.
After starting the HTTP server, add the following to ~/.codex/config.toml:
[mcp_servers.pyghidra-mcp]
url = "http://127.0.0.1:8000/mcp"Restart Codex after saving the configuration. The server name is arbitrary; pyghidra-mcp is the name that will appear in Codex.
Use stdio only when an MCP host cannot use an HTTP endpoint. It starts a server that communicates over its parent process's standard input and output streams.
pyghidra-mcp --transport stdioBecause the server communicates on standard input and output, it will appear to wait without terminal output; that is expected.
docker run -i --rm ghcr.io/clearbluejar/pyghidra-mcp -t stdioThe Docker image starts the Streamable HTTP server by default, so include -t stdio after the image name and use -i for interactive stdio mode.
Use MCPO only when OpenWeb-UI or another consumer specifically requires an OpenAPI proxy. For normal MCP clients, connect directly to the Streamable HTTP endpoint above.
openweb.mp4
uvx mcpo -- pyghidra-mcp /bin/lsuvx mcpo -- docker run -i --rm ghcr.io/clearbluejar/pyghidra-mcp /bin/lsWarning
The MCP community considers this a legacy transport protocol intended for backwards compatibility. Streamable HTTP is the recommended replacement.
SSE transport enables server-to-client streaming with Server-Send Events for client-to-server and server-to-client communication. See the spec for more details.
By default, the server listens on http://127.0.0.1:8000/sse for client connections. Use --host / --port or the MCP_HOST / MCP_PORT environment variables to change the bind address. The server must be running for clients to connect to it.
pyghidra-mcp -t sseBy default, the Python package will run in stdio mode, so you will have to include -t sse.
docker run -p 8000:8000 ghcr.io/clearbluejar/pyghidra-mcp -t sse
This project implementation and design was inspired by these awesome projects:
We believe the future of reverse engineering is agentic, contextual, and scalable.
pyghidra-mcp is a step toward that future—making full Ghidra projects accessible to AI agents and automation pipelines.
We’re actively developing the project and welcome feedback, issues, and contributions.
Note
We love your feedback, bug reports, feature requests, and code.
If you're adding a new tool or integration, here’s the recommended workflow:
- Label your branch with the prefix
feature/to indicate a new capability. - Add your tool using the same style and structure as existing tools in
pyghidra/tools/. - Write an integration test that exercises your tool using a
StdioClientinstance. Place it intests/integration/. - Extend concurrent testing by adding a call to your tool in
tests/integration/test_concurrent_streamable_client.py. - Run make test and make format to ensure your changes pass all tests and conform to linting rules.
This ensures consistency across the codebase and helps us maintain robust, scalable tooling for reverse engineering workflows.
Made with ❤️ by the PyGhidra-MCP Team