Ghidra Headless MCP

Ghidra Headless MCP is a headless Ghidra server that speaks the MCP protocol, letting AI agents do full reverse engineering without opening a GUI: program lifecycle management, disassembly, decompilation, patching, types and symbols, cross-references and script execution, all through tool calls. It is designed to run in the same Docker container as the agent runtime, with no extra sidecar services needed.

Its most concrete selling point is the size of the toolset: 212 tools in 34 groups, from opening programs and managing projects to looking up functions, variables, types, memory, search and graph extraction — basically the whole day-to-day reverse-engineering workflow. Read-only is the default, and changes go through transactions, undo/redo and explicit saving, which is a safety lock on the idea of letting an AI touch binaries. It also ships a command-line client, ghidra_cli, so agents and scripts that don't speak MCP can use the same tools, with batch calls and long-session management supported.

Getting started needs two things: Python 3.11 or later, plus a Ghidra install with the pyghidra runtime. You can still play with it without Ghidra: the project includes a fake backend mode made for CI and development debugging, which runs the whole flow without a Ghidra install. It supports stdio and TCP transports, and Claude Code, Codex and generic MCP clients all have ready-made connection instructions — paste the config from the README and it connects.

Two things to flag: first, this is a local tool with no web demo, so the demo link is the repo page itself, and it suits reverse engineers and security researchers who already use Ghidra and want agents to take over repetitive analysis work. Second, the repo has 196 stars, is GPL-2.0 licensed, and its latest commit was on 16 September 2026, so updates are not frequent — if that matters to you, watch it a while first. Its positioning is clear: not an app for ordinary users, but analysis infrastructure for agents.