ghidra-retro-mcp
Enables AI assistants to perform headless reverse engineering of retro game ROMs using Ghidra, with automated platform detection and triage.
README
Ghidra BizHawk MCP
A unified MCP (Model Context Protocol) server bridging Ghidra's headless static analysis with BizHawk's live emulation — switch between decompiling a ROM and running it on real hardware in the same session.
GBA ROMs: If analyzing Game Boy Advance ROMs, install pudii/gba-ghidra-loader in your Ghidra installation for proper ROM header parsing, mirrored memory regions, and I/O register maps. The loader repository has pre-built
.gpafiles for Ghidra 11.x.
Prerequisites
| Dependency | Version | Required | Notes |
|---|---|---|---|
| Python | >= 3.10 | Yes | Runtime for the MCP server |
| Ghidra | 11.x or 12.x | Yes | Headless or GUI install; GHIDRA_INSTALL_DIR must point here |
| Java (JDK) | >= 17 | Yes | Bundled with Ghidra; needed for JVM bridge |
| pyghidra | >= 3.0 | Yes | Python-to-Ghidra bridge; installed automatically |
| BizHawk (EmuHawk) | Latest stable | No | Only needed for live emulation tools; BIZHAWK_EXE_PATH optional |
| Docker | Latest | No | Only needed for containerized deployment |
Architecture
┌─────────────────────────────────────────────────────────────────┐
│ MCP Client (Claude / Cursor) │
│ sends JSON-RPC over stdin/stdout │
└──────────────┬──────────────────────────────────────┬───────────┘
│ │
▼ ▼
┌──────────────────────────────┐ ┌──────────────────────────────┐
│ ghidra-bizhawk-mcp │ │ BizHawk Lua Bridge │
│ ┌─────────────────────────┐ │ │ ┌────────────────────────┐ │
│ │ MCP Server (server.py)│ │ │ │ bridge.lua │ │
│ │ tool registry, stdio │─┼────┼─>│ memory read/write │ │
│ │ dispatch │ │ │ │ joypad, savestate │ │
│ └───────────┬─────────────┘ │ │ │ frame advance │ │
│ │ │ │ └────────────────────────┘ │
│ ┌───────────▼─────────────┐ │ └──────────────────────────────┘
│ │ GhidraSession │ │
│ │ pyghidra JVM bridge │ │
│ │ decompile, search, │ │
│ │ emulate, fingerprint │ │
│ │ ┌───────────────────┐ │ │
│ │ │ Ghidra JVM (Java) │ │ │
│ │ │ EmulatorHelper │ │ │
│ │ │ SLEIGH / P-code │ │ │
│ │ └───────────────────┘ │ │
│ └─────────────────────────┘ │
└──────────────────────────────┘
All communication is local stdio only — there is no HTTP listener, no TCP socket, and no network exposure.
Security Model
This server communicates exclusively over standard process stdio — there is no HTTP socket, no TCP listener, and no network interface exposed. It is inherently immune to LAN/WAN exposure, SSRF, and unauthenticated API attacks. The only way to interact with it is for an MCP client to launch it as a subprocess and communicate via stdin/stdout.
Hardware & Retro Ecosystem Integration
ghidra-bizhawk-mcp includes native out-of-the-box support for retro-reversing automation pipelines via Ghidra's static analysis, plus live emulation via BizHawk's multi-system emulator. The server bundles:
- Nintendo Entertainment System (NES) via
GhidraNes - Super Nintendo Entertainment System (SNES) via native 65816 memory maps
- Game Boy Advance (GBA) via
gba-ghidra-loader - Nintendo DS (NDS) via
NTRGhidra - Nintendo Switch via
ghidra-switch-loader - PlayStation 1 (PSX) via
ghidra_psx_ldr - Sega Genesis / Mega Drive via native 68000 memory maps
- Sega Master System / Game Gear via
Ghidra-SegaMasterSystem-Loader - Sega Dreamcast via native SuperH4 memory maps
Zero-Input Triage — Worked Example (GBA)
The primary entry point is triage_and_load_retro_rom. Call it with any ROM path and the server handles the rest:
# Auto-detect platform, map language, provision session
triage_and_load_retro_rom(rom_path="/data/game.gba")
# → platform: "Game Boy Advance (GBA)"
# → loader: "GBA ROM Loader"
# → arch: "ARM:LE:32:v4t"
# Decompile the main entry point on the same session
decompile_function(address="0x00001c2c")
# → decompiled C code for the GBA ROM entry routine
# Search for a known pattern (e.g. 32-bit ARM store-multiple)
search_bytes(pattern="09 08 00 01")
# → matching addresses labelled "gba_ram_start"
Execution Chaining Flow
Instead of forcing your AI agent to spend cycles manually identifying architecture maps, register layouts, or memory segments, chain the automated ingestion pipeline:
- Invoke
triage_and_load_retro_romwith a target file path. - The server headlessly parses the binary file structure (
NES\x1a,NTR,NSO0,GBA, SNES title vectors,PS-X EXE,SEGA,TMR SEGA,SEGA ENTERPRISES), binds the matching Ghidra language module (6502:LE:16,ARM:LE:32:v4t,AARCH64:LE:64,65816:LE:24,MIPS:LE:32,68000:BE:32,Z80:16,SuperH4:LE:32), loads standard address memory blocks, and links automated signature cache arrays. - Use the integrated
emulate_sliceoremulate_slice_with_tainttools to analyze localized console loops — no physical console hardware or open GDB networking ports needed.
Triage Tool
| Tool | Description |
|---|---|
triage_and_load_retro_rom |
Reads raw file magic bytes to detect NES, SNES, GBA, NDS, Switch, PSX, Genesis, SMS, or Dreamcast ROMs. Provisions a correctly-language-mapped Ghidra session and auto-restores cached function signatures. Returns platform, loader, architecture tag, and mapped memory blocks. |
Quick Start
1. Install
pip install ghidra-bizhawk-mcp
Or from source:
git clone https://github.com/getanirao/ghidra-bizhawk-mcp.git
cd ghidra-bizhawk-mcp
pip install -e .
2. Set environment
# Required: point to your Ghidra installation
export GHIDRA_INSTALL_DIR=/opt/ghidra_11.2 # Linux / macOS
set GHIDRA_INSTALL_DIR=C:\Program Files\ghidra_11.2 # Windows
# Optional: enable live BizHawk emulation
export BIZHAWK_EXE_PATH=/path/to/EmuHawk.exe
# Optional: run in mock mode (no Ghidra/BizHawk needed)
export MOCK_MODE=1
3. Run
ghidra-bizhawk-mcp
The server listens on stdin/stdout — pipe it to any MCP-compatible client.
Docker
docker build -t ghidra-bizhawk-mcp .
docker run -i --rm -v /path/to/binaries:/data ghidra-bizhawk-mcp
The container bundles JDK 17, Ghidra 11.2, and the server — no host dependencies beyond Docker.
Configuration
Environment Variables
| Variable | Required | Default | Description |
|---|---|---|---|
GHIDRA_INSTALL_DIR |
Yes | — | Path to Ghidra installation (e.g. /opt/ghidra_11.2) |
BIZHAWK_EXE_PATH |
No | — | Path to EmuHawk.exe for live emulation features |
MOCK_MODE |
No | 0 |
Set to 1 to run without Ghidra/BizHawk (for testing/CI) |
Claude Desktop
Add to your claude_desktop_config.json:
{
"mcpServers": {
"ghidra-bizhawk": {
"command": "ghidra-bizhawk-mcp",
"env": {
"GHIDRA_INSTALL_DIR": "/opt/ghidra_11.2"
}
}
}
}
Cursor
Add to your Cursor MCP configuration:
{
"mcpServers": {
"ghidra-bizhawk": {
"command": "ghidra-bizhawk-mcp",
"env": {
"GHIDRA_INSTALL_DIR": "/opt/ghidra_11.2"
}
}
}
}
Tools
Session management
| Tool | Description |
|---|---|
analyze_binary |
Import + analyze a binary, returns a session_id. Reuses the ID if provided, otherwise auto-generates. |
list_sessions |
List all active workspaces with their session IDs, binary paths, and load times. |
close_session |
Close a session and free its Ghidra project resources. |
Most tools accept an optional session_id parameter — omit it to use the most recently loaded session.
Read / Analysis
| Tool | Description |
|---|---|
decompile_function |
Decompile a function by name or address. |
decompile_function_paginated |
Decompile with line_start, line_end, max_tokens (token-budget truncation), and summarize (strips boilerplate locals + collapsing blank lines). Prevents context-window exhaustion. |
get_data_types |
List all data types defined in the program. |
get_cross_references |
Cross-references to/from an address. |
get_call_graph |
Recursive call graph + callers for a function. |
analyze_and_decompile_entrypoints |
Composite — bulk decompile all entry points (program entry, exports, main, _start, etc.) in one call. |
generate_workspace_report |
Produce a Markdown summary of the active workspace — entry points, function count, custom symbols, recovered structures, renamed functions, comments. Replaces a GUI CodeBrowser window. |
Write / Mutation
| Tool | Description |
|---|---|
rename_symbol |
Rename a function or label. Stored in the Ghidra project DB. |
add_comment |
Attach a comment (plate, pre, post, eol, repeatable). |
create_struct |
Create a custom structured data type from a JSON member layout [{offset, name, type}, ...]. Offsets are optional. |
retype_variable |
Re-type a local variable or function parameter (e.g. undefined4* → MyStruct*). |
Assembly-level
| Tool | Description |
|---|---|
disassemble_range |
Disassemble N raw instructions at an address — returns mnemonic, operands, hex bytes, and length for precise lower-level inspection. |
get_listing_range |
Raw hex + ASCII dump for a byte range, equivalent to Ghidra's Listing panel. Complements disassemble_range for data regions. |
Byte-sequence search
| Tool | Description |
|---|---|
search_bytes |
Search the entire binary for a hex byte pattern (e.g. 09 08 00 01 or F86D0003). Returns matching addresses with context bytes and any string label at the hit. |
Binary diffing
| Tool | Description |
|---|---|
diff_binaries |
Compare two loaded sessions by function name and body size. Returns functions unique to each side and changed functions. |
Workspace Sessions
Each analyze_binary call creates a named session. Sessions keep their Ghidra project open independently, so multiple binaries can be loaded concurrently:
# Load two binaries into separate sessions
s1 = analyze_binary(binary_path="/bin/a.out") # auto session_id
s2 = analyze_binary(binary_path="/bin/b.out", session_id="my_session")
# Operate on a specific session
decompile_function(function_name="main", session_id=s1.session_id)
# Diff them
diff_binaries(session_a=s1.session_id, session_b="my_session")
Deployment
Docker (multi-user / CI)
docker build -t ghidra-bizhawk-mcp .
# Run as an MCP subprocess
docker run -i --rm \
-v /data/binaries:/data \
ghidra-bizhawk-mcp \
--ghidra-dir /opt/ghidra
The Dockerfile bundles Ghidra 11.2 and JDK 17 in a slim Python 3.11 image. Bind-mount your binaries directory at runtime.
MCP Bundle (MCPB — Claude Desktop / Smithery)
Package as a portable .mcpb bundle for one-click install in Claude Desktop or publishing on Smithery.
Prerequisites: Install the MCPB CLI:
npm install -g @anthropic-ai/mcpb
Build the bundle:
# From the repo root
scripts/build-mcpb.ps1
Or manually with mcpb:
mcpb pack
The output ghidra-bizhawk-mcp.mcpb wraps the server with a manifest.json that prompts for GHIDRA_INSTALL_DIR (required) and optionally BIZHAWK_EXE_PATH at install time — no manual JSON editing.
Publishing to Smithery:
smithery mcp publish ./dist/ghidra-bizhawk-mcp.mcpb -n getanirao/ghidra-bizhawk-mcp
P-code micro-emulation
| Tool | Description |
|---|---|
emulate_slice |
Headlessly execute N instructions. Seed register state and get a step-by-step trace of register mutations. |
emulate_slice_with_taint |
Same as emulate_slice but with automated taint tracking — specify a taint register (e.g. r0) and the tool flags exactly when its value is modified or propagates to other registers. |
emulate_slice_with_breakpoints |
Execute until a condition is met or the count expires. Condition syntax: R0==0, R1>0xFF, R2!=R3, PC==0x1234. Stops before or after the matching instruction. |
All run inside the pyhidra process via Ghidra's EmulatorHelper — no GDB/LLDB, no network ports, no debugger stubs. Works on ARM, x86, MIPS, and any Ghidra-supported architecture.
Worked example — breaking on a register condition
Suppose you're reversing a GBA ROM and want to find the first time r0 becomes zero inside a loop at 0x08000100:
# Step until r0 == 0, stop before the matching instruction
result = emulate_slice_with_breakpoints(
session_id="gba_v1",
start_address="0x08000100",
max_instructions=5000,
stop_condition="R0==0",
stop_mode="before"
)
# result.exit_reason → "R0==0"
# result.instructions_executed → 312
# result.trace → [step 311: r0 goes 4→2, step 312: r0 goes 2→0]
# Check if a specific address was reached after a branch
result = emulate_slice_with_breakpoints(
session_id="gba_v1",
start_address="0x08000100",
max_instructions=5000,
stop_condition="PC==0x08001234"
)
# result.exit_reason → "PC==0x08001234"
# Use inequalities to catch bounds checks
result = emulate_slice_with_breakpoints(
session_id="gba_v1",
start_address="0x08000100",
max_instructions=5000,
stop_condition="R1>0xFF"
)
# result.exit_reason → "R1>0xFF"
# result.last_step["r1"] → 0x100
This is especially powerful for identifying copy-loop bounds (R3 >= R4), null-pointer paths (R0==0), or switch-table targets (PC==0x).
Function fingerprinting / signature transfer
| Tool | Description |
|---|---|
calculate_function_fingerprint |
Generate a structural hash for a function (vars, params, body size, branches, called funcs, embedded strings, numeric constants). Survives compiler reordering. |
export_signature_map |
Build a complete {hash → name} map for every function in the current binary. Save this JSON to reuse across versions. |
apply_signature_map |
Pass a previously exported signature map; the server sweeps the binary and renames every matching function automatically. |
Persistent signature stash (server-side cache)
| Tool | Description |
|---|---|
save_active_binary_signature |
Fingerprint all functions and stash the map under a lineage_group_id (e.g. "my_firmware_v1"). Stored in ~/.ghidra_bizhawk_mcp/signatures/ — no JSON files to manage. |
auto_restore_signatures_from_stash |
Load a stashed map by lineage_group_id and auto-rename every matching function. |
auto_stash_current_binary |
Zero-input auto-stash — hashes the binary's first 4 KB, saves a map under that hash. Just analyze and call. |
auto_restore_current_binary |
Zero-input auto-restore — hashes the binary, looks up a previous stash, renames matches. No group ID needed. |
list_stashed_signature_groups |
List all stashed groups currently in the local cache. |
Workflow — fully automated persistence:
# Analyze v1 — stashes automatically under binary content hash
s1 = analyze_binary(binary_path="/bin/v1.bin")
auto_stash_current_binary(session_id=s1.session_id)
# Later, analyze v2 — restores automatically
s2 = analyze_binary(binary_path="/bin/v2.bin")
auto_restore_current_binary(session_id=s2.session_id)
# → 142 functions renamed, zero manual JSON handling
Demo

Claude Desktop: "Decompile the entry point of this GBA ROM and trace r0 propagation" — the server auto-detects the ARMv4t language, provisions a session, and returns decompiled C + taint trace.

Console output from triage_and_load_retro_rom detecting a PlayStation 1 executable (PS-X EXE magic), mapping MIPS:LE:32, and auto-restoring cached signatures.
Try these prompts
After configuring your MCP client (see Configuration), ask your AI agent:
- "Load this GBA ROM and decompile the entry point."
- "What functions call 0x8001234 in this NDS binary?"
- "Triage this PSX EXE and trace r0 through the first 20 instructions."
- "Diff the two sessions I have open and show me changed functions."
Project Structure
ghidra-bizhawk-mcp/
├── Dockerfile
├── pyproject.toml
├── README.md
└── src/ghidra_bizhawk_mcp/
├── __init__.py
├── server.py # MCP server, tool registry, stdio transport
├── ghidra_bridge.py # GhidraSession — pyghidra wrapper, all Ghidra logic
├── lua/
│ └── bridge.lua # BizHawk-side Lua bridge for live emulation
└── tools/
├── __init__.py
├── bizhawk_bridge.py # TCP server bridging MCP ↔ BizHawk
└── ...
How it works
pyhidra.start()boots Ghidra's JVM once at server startup- Each
analyze_binarycall opens a new Ghidra project in its own named session - Read/write tools route to the requested session via
session_id(or the active default) - Write tools apply changes directly to the Ghidra program database
- Sessions persist until explicitly closed — enabling multi-binary workflows and diffing
<!-- mcp-name: io.github.getanirao/ghidra-bizhawk-mcp -->
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.