中文版

leanrea（GitHub：https://github.com/aicodeking/leanrea ，2026-10-17 创建，今晨抓取时已 4,047★）是一个 8KB 的技能文件，做的是 REA（Reverse Engineer Anything）做的事——只不过不要 MCP 服务器、不要 207 个工具、不要任何安装：它直接告诉 Agent 对每种目标用哪条现成命令最快（asar、otool、nm、go version -m、jadx、apktool、Ghidra、tshark、binwalk、sqlite3……），并给出即贴即用的侦察命令与 REA 代码里藏着的坑，Agent 自己跑真实工具。MIT 开源。

作者的实测对比很直白：REA 是 25MB 的 npm 包 + MCP 服务器，207 个工具定义占 3.35MB，加载全部工具后“回复 OK”就要 400.7k token（约 $1.60）；leanrea 只是一个 SKILL.md，零新增工具，同样一步只要 34.5k token（约 $0.09）。在 4 个逆向任务 × 3 次重复的 Claude Code 实测里，leanrea 比 REA 快 38%、便宜 35%、轮次少 56%、token 少 60%，得分 74/75，与 REA 的 75/75 基本持平（唯一丢的半分是把旧版 IPC 通道当成了新版）。

安装一条命令：npx skills add aicodeking/leanrea，按提示选 Agent（Claude Code、Codex、Cursor、Gemini CLI、OpenCode 等）；也可手动把 leanrea 文件夹拷进 ~/.claude/skills。用法就是 /leanrea 加一句要查的问题。

为什么值得了解

• 对偶尔要逆向看代码的人，这是“少即是多”的现成例子：一个 8KB 文本文件，顶掉了一个 25MB 的 MCP 服务器，效果几乎一样
• 它点破了一个 Agent 工具设计的通病：工具定义本身就是上下文税，207 个工具光加载就要 400k token——能用现成命令行解决的事，别急着包一层 MCP
• 但要注意实测样本不大（每组 3 次、一个模型、一台 Mac），且作者也承认：不装任何技能、只靠模型本身，也能拿到 67/75 分——技能的增益没有数字看起来那么夸张

English Version

leanrea (GitHub: https://github.com/aicodeking/leanrea , created 2026-10-17, 4,047 stars at this morning’s fetch) is a single 8 KB skill file that does what REA (Reverse Engineer Anything) does — with no MCP server, no 207 tools, no setup: it tells the agent which existing command is fastest for each kind of target (asar, otool, nm, go version -m, jadx, apktool, Ghidra, tshark, binwalk, sqlite3 and more), gives ready-to-paste recon commands, and lists the traps REA’s code handles for you, while the agent runs the real tools itself. MIT licensed.

The author’s benchmark is blunt: REA is a 25 MB npm package plus an MCP server, its 207 tool definitions take 3.35 MB, and loading all of them costs 400.7k tokens (about $1.60) just to say “Reply OK”; leanrea is one SKILL.md, adds zero tools, and the same step costs 34.5k tokens (about $0.09). Across 4 reverse-engineering tasks × 3 repetitions in headless Claude Code, leanrea was 38% faster, 35% cheaper, 56% fewer turns and 60% fewer tokens than REA, scoring 74/75 against REA’s 75/75 (its one partial miss named an older IPC channel instead of the v2 one).

Install with one command: npx skills add aicodeking/leanrea, then pick your agents (Claude Code, Codex, Cursor, Gemini CLI, OpenCode and others); or copy the leanrea folder into ~/.claude/skills by hand. Then use it as /leanrea followed by your question.

Why it’s worth knowing

• For anyone who occasionally reverse-engineers code, this is a ready-made example of “less is more”: one 8 KB text file replaces a 25 MB MCP server at nearly the same result
• It exposes a common flaw in agent tool design: tool definitions are themselves a context tax — 207 tools cost 400k tokens just to load — so if a plain shell command can do the job, don’t rush to wrap it in MCP
• Mind the sample size, though (3 repetitions per cell, one model, one Mac), and the author admits the plain model alone scores 67/75 — the skill’s gain is smaller than the headline numbers suggest