Comment by fullstackchris
10 hours ago
> The reason I think MCP may be a one-year wonder is the stratospheric growth of coding agents. It appears that the best possible tool for any situation is Bash—if your agent can run arbitrary shell commands, it can do anything that can be done by typing commands into a terminal.
I push back strongly from this. In the case of the solo, one-machine coder, this is likely the case - if you're exposing workflows or fixed tools to customers / collegues / the web at large via API or similar, then MCP is still the best way to expose it IMO.
Think about a GitHub or Jira MCP server - commandline alone they are sure to make mistakes with REST requests, API schema etc. With MCP the proper known commands are already baked in. Remember always that LLMs will be better with natural language than code.
The solution to that is Anthropic's Skills.
Create a folder called skills/how-to-use-jira
Add several Bash scripts with the right curl commands to perform specific actions
Add a SKILL.md file with some instructions in how to use those scripts
You've effectively flattened that MCP server into some Markdown and Bash, only the thing you have now is more flexible (the coding agent can adapt those examples to cover new things you hadn't thought to tell it) and much more context-efficient (it only reads the Markdown the first time you ask it to do something with JIRA).
But that moves the burden of maintenance from the provider of the service to its users (and/or partially to intermediary in form of "skills registry" of sorts, which apparently is a thing now).
So maybe a hybrid approach would make more sense? Something like /.well-known/skills/README.md exposed and owned by the providers?
That is assuming that the whole idea of "skills" makes sense in practice.
Yeah that's true, skill distribution isn't a solved problem yet - MCPs have a URL, which is a great way of making them available for people to start using without extra steps.