MCP server exposing 180+ reverse-engineering tools that aggregate Ghidra, IDA, x64dbg, Frida and 56 other backends with native binary parsers and a 4,200-skill knowledge base.
230+ MCP tools · 60+ RE/debug backends · 9 host-side plugins · 6,000+ skills · 170+ agents · 160+ subagents · 190+ plugins · 3,000+ prompts · multi-agent mesh · native x86-64 + ARM64 emulators
v5.0.0 released. 19 GiB → 3.3 GiB working tree (83% smaller via gzipped source archives), + four feature waves: host plugin build systems, Frida bridge + dynamic analysis, ARM64 emulator + decompile cache + chunking, bindiff / ROP / game-engine metadata / MCP sampling / SVG visualization. See CHANGELOG.
Quickstart · Architecture · Comparison · Agent Mesh · MCP Hub · Ethics
Reverse engineering a modern application involves five static tools (Ghidra, IDA, Binary Ninja, radare2, Hopper), three debuggers (x64dbg, WinDbg, GDB/LLDB), a hex-editor, a memory scanner, an instrumentation framework, and a stack of format parsers (PE, ELF, Mach-O, DEX, Minidump, WASM, .NET, ASAR, GGUF, BEAM, Go). Switching between them costs hours. Teaching a model to drive even one of them well costs more.
This server:
smart_decompile picks IDA when running, falls back to Ghidra, then Binary Ninja, then radare2.decompile setup, decompile analyze <path>, decompile capabilities, decompile doctor) that registers the MCP with Claude Code / Cursor / Codex / Windsurf / Gemini and returns structured findings with evidence and unknowns.If you're on security-research, binary-analysis, defensive-engineering, or SDK work — this collapses a day of tool-shuffling into one MCP connection.
Needs Node.js 24+ for the MCP server (how to get it). Pick one:
# Linux / macOS / FreeBSD / Termux
curl -fsSL https://raw.githubusercontent.com/1llu50ry/Decompile/main/install.sh | sh
# Windows (PowerShell)
irm https://raw.githubusercontent.com/1llu50ry/Decompile/main/install.ps1 | iex
# npx (any OS, nothing kept but the RE tools)
npx -y decompile@latest setup -i --npx
Each one installs everything and wires it up:
setup -i, 42 agents);~/.decompile/tools (~1.5 GB): Ghidra + Java 21, radare2, UPX, x64dbg (Windows),
FakePDB, and a Python env with frida, angr, capstone, keystone, unicorn, lief, r2pipe, pyghidra,
volatility3, yara, mitmproxy, ROPgadget — then env.json hooks them into the server;--minimal (-Minimal on Windows) installs only the core; add the RE programs later with
decompile retools install. Other flags: --dry-run, --version v0.0.1, --method npm|npx,
--install-node, --no-setup, --uninstall — see the Installation wiki.
Check the result with decompile doctor; manage RE programs with decompile retools status | install | env.
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup
By default it configures Claude Code. Pass -i for an interactive
picker over 42 supported agents (Claude Code / Desktop, Codex, Cursor,
VS Code, Zed, Windsurf, Gemini CLI, Grok Build, Qwen Code, Kimi,
DeepSeek Harness, LM Studio, Cline in every host, Copilot CLI, Goose,
and 25+ more). The workflow-instructions skill is installed alongside
the registration. Restart your agent and you're done.
Variations:
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup -i # interactive picker (42 agents)
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup --all-detected # every detected agent
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup --client claude_code # one specific client
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup --client codex --client cursor --client dsh
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup --dry-run # preview without writing
npx -y https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz setup --npx # cache-safe (npx-based) registration
npm install -g https://github.com/1llu50ry/Decompile/releases/latest/download/decompile.tgz && decompile --help # install the command permanently
# clone + build (local dev)
git clone https://github.com/1llu50ry/decompile && cd decompile
npm install && npx tsc
npm test # 20/20 should pass
# run the MCP server over HTTP (Docker-friendly)
docker compose up
# drive it from the CLI (no MCP client needed)
decompile analyze path/to/binary # structured findings JSON
decompile capabilities # list every tool + skill
decompile doctor # runtime + DB + registrations
decompile providers # list configured RE backends
decompile guidance # print the workflow instructions
Understand how feature X works in that app, show me the evidence,
and port it to my project.
Replace X and that app with your actual target; the server returns
structured findings with evidence tool-backed and unknowns called out.
Point your MCP client at it, then ask any of: