Replacing the Semantic Scholar MCP with a Skill?

Thu, 30 Apr 2026 00:00:00 +0000

In the previous post, the Semantic Scholar MCP beat a vanilla agent on a reference extraction task with 100% recall vs. 85%, 3x fewer tool calls, and zero side effects. But since everybody’s buzzing how MCP is dead, skill is all you really need, yada yada yada… in this post I’ll show is this really true. Can a skill actually do the same thing, but cheaper?

What Changed

So, I built a /scholar skill. It’s extremely simple: three Python scripts calling the same Semantic Scholar API, using stdlib urllib and running via uv. No async framework, no rate limiter, no circuit breaker, no nothing. The agent (with frontier model that it’s using, of course) is more than capable enough to retry if a call fails.

Skills on Ante Kapetanovic

Replacing the Semantic Scholar MCP with a Skill?

What Changed