Reverse Engineering Without the MCP Bloat

If you've had a coding agent poke at a binary, you've probably seen the context bill. REA (Reverse Engineer Anything) is an MCP server that adds 207 reverse-engineering tools, with 3.35 MB of tool definitions. leanrea replaces all of that with one 8 KB SKILL.md. It adds zero tools. It tells the agent which existing CLI tool to use for each kind of target, and the agent runs those through Bash. The README measures a "reply OK" prompt with all tools loaded: 34.5k tokens for leanrea, 400.7k for REA.

How it works

The skill is a router. The agent first checks the target with file and its magic bytes, then picks a row. Electron apps get npx @electron/asar, rg and npx @electron/fuses. Mach-O gets otool, xcrun dyld_info, nm and codesign. Go binaries get go version -m. The table also covers ELF/PE, .NET, Android APKs, WASM, SQLite, pcap/HAR, firmware and running processes (lldb, dtrace, frida).

The useful part is the pitfalls list, and the author credits most of it to reading REA's code. Bundlers hide code in .webpack/ and .vite/, so search with rg -uu. Don't cat minified files, slice around a match instead. Bytenode .jsc files can't be decompiled, but Hermes bytecode can. Every claim gets labeled as seen, inferred or unknown.

Raw commands are also faster. Per the README, a Mach-O overview takes 3.32 s through REA and 0.04 s with otool, dyld_info and nm.

Try it

npx skills add aicodeking/leanrea
npx skills add aicodeking/leanrea -g -a '*' -y

# or manually
git clone https://github.com/aicodeking/leanrea
mkdir -p ~/.claude/skills && cp -R leanrea/leanrea ~/.claude/skills/

/leanrea How does the search feature in Foo.app work? Show the evidence.

If your agent doesn't support skills, paste SKILL.md into its rules or system prompt.

My take

I like the premise. Most of what an MCP wrapper does for CLI tools, a good page of notes can do too, and here it's been benchmarked. In Claude Code with Sonnet, leanrea came in 38% faster and 35% cheaper than REA, scoring 74/75 against 75/75.

The author is upfront about the limits. That's 3 reps on one model and one machine. REA ran in its best case, with deferred tool loading. Ghidra, Hopper and jadx weren't installed. Plain Sonnet with no skill already scored 67/75, so the gain over no skill (about 19% faster, 32% fewer turns) is real but smaller. Since the skill only routes to tools, you still have to install those tools yourself.

If you use agents to dig into Electron apps or macOS binaries, try this. It's one small markdown file, and it's MIT licensed.