TestMu AI, the company the catalog still knows as LambdaTest, has published the first skills-versus-MCP piece that measures instead of analogizes. In MCP vs Agent Skills: What Each Is For and When to Use Both, Anubhav Singhmaar takes the 71 skills in the company’s open source library and counts what an agent carries at each level of progressive disclosure. The always-resident metadata across all 71 skills comes to 33,016 characters. The full skill bodies come to 402,822. That is a 12.2x multiple between what the agent always holds and what exists to be loaded, and the median skill is 383 characters at level one against 5,259 characters and 174 lines at level two. The argument that follows is short: “Access problems are MCP problems. Method problems are skill problems, and they are far more common on teams that have already connected their tooling.”
The numbers are the vendor’s, measured on the vendor’s own library, and a 71-skill corpus written by one company will not be representative of anything but that company. What survives is the shape. Level-one metadata is cheap enough to preload everywhere, level-two bodies are not, and the post is candid that “loading all 71 bodies at once would be the wrong design.” The practical advice holds up on its own: a skill “is the cheaper experiment either way, since it is one markdown file with two required fields and it goes through the same review as your code.” Start there, add MCP when the agent needs live data it cannot otherwise reach. The company distributes its Kane CLI browser automation the same way, as a skill rather than an MCP server, which is a stronger commitment than a blog post.
This is a story where the catalog agrees with the provider. On the LambdaTest provider page both the agent skills dimension and the MCP server dimension are lit, which puts the company in the small set of providers that ship both layers the post is describing. The catalog lists 20 API pages beneath the provider. The flaky-test example in the post, where a skill triggers on a description match and then calls MCP tools for live run data, would land on surfaces like the LambdaTest Session Logs (V2) API and the LambdaTest Test API. The agentic access profile maps 52 operations, 20 of them acting and 3 flagged human-in-the-loop, which is the access layer the skills are meant to sit on top of.
The Kin Score is 45.9, in the developing band. Discoverability carries it at 68.5 and contract quality follows at 61.2, with operational transparency at 42.1. Contract governance sits at 9.8, and the well-known catalog dimension is unlit, which is the one gap the story itself exposes. The 71 skills live in a GitHub repository the post links to, not at a well-known location on the company’s own domain where an agent could find them without being told. A company that just measured how cheaply metadata can be preloaded has not yet published the one file that would let a stranger’s agent preload it.