Yash Patil (@yvshpvtil) on X

What it claims

Yash Patil's X post reports that leanrea, a single 8 KB skill file, matched REA, a 207-tool MCP server for reverse engineering, on the leanrea author's own test. It scored 74/75 against REA's 75/75, ran 38% faster and cost 35% less. Patil credits the screenshot to @aicodeking. Patil's view is that "many MCP servers are instructions in disguise." The post adds one fact the catalog summary lacks: adoption. Citing ossradar.net, Patil says the repository had 392 GitHub stars on October 17 and 1,634 at posting time. The post does not give a year.

How it works

The post describes no internals beyond the summary. leanrea runs no server. It is a text file that tells a coding agent which plain CLI tools to call for reverse-engineering tasks.

A reply in the thread offers a reason for the speed and cost gap. It is the replier's opinion, not a measured result. The argument is that the schemas for 207 tools sit in the model's context and add cost to every call. A text file avoids that overhead.

Limits

The leanrea author lists these caveats: 3 runs per task, one model, one machine. The author also notes that Claude with no skill at all was already decent. The post gives no number for that no-skill baseline.

The same reply raises an open question. It asks whether the one lost point came from a task where structured MCP tool output beat parsing CLI text. It asks for results split by task type, and the post provides none.

The post also leaves out several things a reader would need to try the tool:

• a link to the repository
• the list of CLI tools the file names
• install steps
• a licence
• the task set behind the 75-point score

Anyone considering adoption has to check those in the repository itself.