Which Ghidra MCP server should I use?
A practical comparison of six Ghidra MCP servers: how each connects, how to install it, licence, last commit and who each one suits.
If you want the simplest start with Claude Code and a Ghidra window you already have open, use LaurieWired's GhidraMCP. If you want to run Ghidra without a window, for scripts or many binaries, use pyghidra-mcp, or ReVa if you want a Ghidra extension with a headless mode. Use bethington's server only if you want the largest tool set and accept a heavier install, and use REA if your targets are not only native code.
All of this is as of 21 November 2026. These projects change weekly, so check each README before installing. I read each repository's README and licence file and took the last commit date from the repository's git history. The repository pages did not show star counts to my fetcher, so I give none rather than guess.
The comparison
| Server | Connects to | Transport (per README) | Install | Licence | Last commit |
|---|---|---|---|---|---|
| LaurieWired/GhidraMCP | Ghidra GUI plugin | stdio via uvx, or streamable HTTP | Extension zip plus uvx | Apache-2.0 | 10 Oct 2026 |
| GhidraPluginProject/ghidra-mcp | Ghidra GUI or headless | Not checked in detail | Build with Maven | Apache-2.0 | 27 Mar 2026 |
| bethington/ghidra-mcp | Ghidra GUI plugin or headless | Python bridge (stdio), HTTP transports | Release zip plus bridge wheel | Apache-2.0 | 20 Nov 2026 |
| clearbluejar/pyghidra-mcp | Headless pyghidra, or --gui | Streamable HTTP (recommended), stdio, legacy SSE | Python package | Apache-2.0 | 6 Nov 2026 |
| cyberkaida/reverse-engineering-assistant (ReVa) | Ghidra extension, UI or headless | Streamable HTTP on port 8080 | Extension for your Ghidra version | Apache-2.0 | 17 Nov 2026 |
| morluto/rea (REA) | Ghidra, Hopper, IDA or Kuna, by target | MCP, set up with npx rea-agents setup | npm | MIT | 21 Nov 2026 |
LaurieWired/GhidraMCP
The best known of the group. Its README describes an MCP server plus a Ghidra plugin that decompiles, renames methods and data, and lists methods, classes, imports and exports. The plugin runs an HTTP server inside Ghidra (port 8080 by default), and the Python MCP server talks to it. The README pins Ghidra 12.1.4 and uses uvx in the client config, with a streamable-HTTP mode for clients that want a URL.
It works on whatever program is open in the Ghidra window, so you need the GUI running. It suits a first session, a CrackMe or a single binary. My step-by-step is in Set up GhidraMCP with Claude Code. Tool page: GhidraMCP.
GhidraPluginProject/ghidra-mcp
Its README describes a Ghidra MCP server with 193 tools, GUI and headless modes, and Docker support, at version 4.3.0 against Ghidra 12.0.3. Install is a Maven build. The last commit I found was on 27 March 2026, much older than the others here. Its README reads very like an earlier version of bethington's. I did not check how the two repositories are related. Unless you have a reason to use it, I would pick one of the newer servers.
bethington/ghidra-mcp
The README claims 219 tools, including P-code emulation and live debugger integration, plus Ghidra Server support and a headless mode. Install is a release extension zip for Ghidra 12.2 plus a Python bridge wheel installed with uv tool install, then you point your client at the bridge-mcp-ghidra command by absolute path. The plugin answers on port 8089. The README says tagged releases are stable and main and dev are beta or experimental. Its last commit was 20 November 2026 (release 7.1.0). It suits people who want one server with many write tools and are happy to read a long README. A larger tool list also means more for the model to choose between.
clearbluejar/pyghidra-mcp
A command-line server built on pyghidra. The README calls it headless-first, with a --gui mode that shares live state with a running Ghidra. It recommends streamable HTTP, keeps stdio for hosts that cannot use HTTP, and lists SSE as legacy. It can analyse every binary in a Ghidra project, starts serving while analysis runs in the background, and has semantic search over decompiled code. The README marks it as beta. Last commit 6 November 2026. It suits scripts, CI and batches of binaries, and anyone who prefers a pip-style install to building a Ghidra extension.
cyberkaida/reverse-engineering-assistant (ReVa)
A Ghidra extension that serves MCP over streamable HTTP, port 8080 by default. It has an assistant mode inside the Ghidra UI and a headless mode where ReVa starts Ghidra and temporary projects for you. Its README says Claude Code is the recommended client and gives claude mcp add --scope user --transport http ReVa -- http://localhost:8080/mcp/message. It supports Ghidra 12.0 and above, and you must download the release that matches your exact Ghidra version (Ghidra 12.2 and later need Java 25). The design goal is many small tools with guided output to limit context use. Last commit 17 November 2026. It suits long sessions on large binaries or firmware.
morluto/rea (REA)
REA is not a Ghidra server first. Its README calls it one MCP for binaries, applications and runtime behaviour, and says native analysis can use an existing Hopper, Ghidra, IDA or Kuna install depending on the target. Setup is npx rea-agents setup. It is MIT licensed and its last commit was 21 November 2026. It suits you if you also look at JavaScript apps, Electron, Android or runtime behaviour, not just one native binary. If you only need Ghidra, a dedicated Ghidra server is simpler. Tool page: REA. A wider comparison including IDA and Binary Ninja is in REA vs GhidraMCP vs IDA Pro MCP.
How to choose
- First time, one binary, Ghidra open on your screen: GhidraMCP.
- No GUI, scripts, CI or many binaries: pyghidra-mcp, or ReVa in headless mode.
- Big binary or firmware and long sessions: ReVa.
- Maximum tool count, emulation and debugger hooks: bethington's server.
- Mixed targets beyond native code: REA.
Whichever you pick, Ghidra itself is free and the servers above are open source, so you can try two on the same binary. Many more Ghidra MCP servers exist than the six here. I could not fetch a published count, so I will not quote one. Names also overlap ("GhidraMCP" is used by several repositories), so check the owner in the URL.
