Yeah that still fails a ton. A skill will mess up the call, not know how to handle errors correctly, not flag an error as an error (false positive), hallucinate, require custom input/output formats, require custom code to validate, won't handle authentication the same way, you'll need a unique skill for each endpoint, you need to version the skill, and of course there needs to be an endpoint, as well as all of the tools needed on the client machine to process everything, hoping of course that they're the right version.

Compare that to MCP. One standard tool call. Works on every agent. Everything built in. Determinism. Reliability. Simplicity. It just works.

Skills are non-deterministic balls of mud. MCP are deterministic standard tool calls. MCP solves all the problems. A skill can't do what MCP does.

In terms of "using standards" to solve the problems, tell me 3 different APIs (from different companies) that have identical input format, identical output format, identical request URI format, identical headers, identical methods, identical auth. You won't find them, because everyone implements their API differently. You can use 50 different standards 50 different ways, and still not solve the problems mentioned above. MCP is one standard that works one way and actually solves all the problems.

And this is just addressing the client-side problems. There's a whole world of server side (logging, tracing, auth, execution, etc) that's not standard either, unless you use MCP.