An 8 KB skill file beats a 207-tool MCP server at its own job

What it does

leanrea is one SKILL.md file — 8 KB of plain text — that teaches an agent how to reverse-engineer binaries, Electron apps, Go builds, and more using the tools already on the machine: otool, nm, asar, go version -m, rg, and friends. No MCP server, no tool definitions, no install beyond dropping the file in place. The agent reads the advice and runs real commands via Bash.

The interesting bit

The pitch is that MCP tool layers can be pure overhead: REA wraps commands the agent can already run, and the wrapper costs time and tokens. The benchmark backs this up — REA’s Electron analysis took 47.9 s and produced 140 MB of JSON where asar extract plus rg -uu took 4.8 s and 24 KB. The skill is essentially distilled gotchas (mostly cribbed from reading REA’s own code, to the author’s credit) plus a router that says which command is fastest for each target type.

Key highlights

• 38% less wall time and 35% lower cost than REA at the same benchmark accuracy (74/75 vs 75/75)
• Zero tools added to the agent’s context — 34.5k tokens vs 400.7k with all REA tools loaded
• Router covers Electron/asar, Mach-O, ELF/PE, Go buildinfo, WASM, APKs, firmware, pcap, and more
• Includes “first moves”: single paste-and-run commands that inventory a whole binary in one call
• Honest methodology: raw per-run metrics published, answer keys versioned to the exact targets

Caveats

• Three reps per cell, one model, one machine — thin statistical ice
• REA was tested in its best case (deferred tool loading); agents that load everything pay the full 400k tokens
• The plain model is already decent at this: the gain over no skill at all (~19% faster) is much smaller than the gain over REA

Verdict

Worth a look if you use agents for reverse-engineering and resent token bloat — it’s also a neat case study in “sometimes a well-written note beats an integration.” If your agent can’t run shell commands, this does nothing for you.