GhidrAssistMCP – AI‑enabled bridge for Ghidra

What it is

GhidrAssistMCP is a Ghidra extension that runs a Model Context Protocol (MCP) server inside the Ghidra application. The server exposes Ghidra’s reverse‑engineering functions (listing functions, getting disassembly, editing comments, running analysis, etc.) through a clean, HTTP‑based API that AI assistants or other automation tools can call.

Why it matters for AI/ML

• Standardised API – By speaking MCP, external LLM‑based assistants (e.g., the companion GhidrAssist client) can ask Ghidra questions and receive structured JSON answers without needing to embed Java code.
• Tool catalog – 100 explicit “tools” are defined with input schemas and read/write annotations, making it easy for an LLM to discover what operations are available and how to call them.
• Async & caching – Long‑running analyses run as background jobs and results are cached, which is useful when an LLM iteratively refines its queries.
• Multi‑program & headless support – The server works both in the GUI and in Ghidra’s analyzeHeadless mode, so AI pipelines can run fully automated analyses on many binaries.

Core features (as described in the README)

Feature	What it does
MCP Server Integration	Implements the official MCP Java SDK (v2.0.1) and exposes two HTTP transports – Server‑Sent Events (SSE) and a streamable JSON endpoint.
Explicit tool catalog	100 named tools (93 enabled by default) covering project management, program inspection, disassembly, decompilation, comments, types, analysis options, and more.
Resource templates & prompts	Pre‑defined MCP resources for program metadata, functions, strings, imports/exports, and segments; plus 7 ready‑made analysis prompts for common RE tasks.
Result caching & async tasks	Repeated queries hit a cache; long‑running operations return a task ID that can be polled or cancelled.
Multi‑program & focus awareness	Handles several open programs, automatically uses the program window that has focus, and lets clients specify a program_name.
Shared runtime & UI	One server instance is owned by the Ghidra project manager; every CodeBrowser window gets a view that shares configuration and logs.
Headless mode	Can be started from analyzeHeadless so AI pipelines can interact with Ghidra without a GUI.
Configurable UI	A Ghidra window (Window → GhidrAssistMCP) lets users enable/disable tools, change host/port, and watch live request/response logs.

How you would use it

1. Install the extension (ZIP) via File → Install Extensions in Ghidra (requires Ghidra 12.0/12.1 and Java 25+).
2. Start Ghidra – the MCP server launches automatically when the project manager opens.
3. Configure host/port if needed and enable the tools you want.
4. Call a tool from any MCP‑compatible client (e.g., the companion GhidrAssist UI or a custom script):

   {"method":"tools/call","params":{"name":"list_functions","arguments":{"program_name":"/binaries/example"}}}
   

   The server replies with a JSON list of functions.
5. Run headless – use Ghidra’s analyzeHeadless launcher with a pre‑script that starts the server, then let an external AI service query the endpoints (/mcp or /sse).

Who might benefit?

• Security researchers building LLM‑driven binary analysis assistants.
• Automation pipelines that need program‑level information (functions, CFGs, decompilation) without embedding Ghidra’s Java API.
• Educational tools that let students explore reverse engineering via a web‑based UI powered by an LLM.

Limitations & considerations (from the docs)

• The server runs on a single port (default 8080); the port must be free.
• Some powerful operations (import_file, export_program, script execution) are disabled by default for safety.
• The extension depends on the MCP Java SDK 2.0.1; newer MCP revisions may need an updated SDK.
• In headless mode the server stays alive only while the JVM is running; a completion_file can be used to shut it down cleanly.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.