Usefulness of MCPs

Where do MCPs break down?

After building a couple MCP servers I've started asking myself: what does this actually provide?

A friend of mine asked his boss if they could add Github MCP to their "allow list" but his company didn't enable it (for one reason or another). He thought to himself "I'd love for my Claude to have access to reviews and issues across our projects". How did he get around this? He just told Claude to use the gh cli and bam, he got the information he wanted. With the right .md files, LLMs should be able to read a cli --man page and know exactly what to do. Where do these provide value again?

What about APIs! You can call your database through the LLM so it can make changes through the DB (yikes?). If that's the goal, why not have the LLM use curl or something? Better yet, write a database cli tool for it to interface with and hard code "cmd != 'DROP DATABASE database_name;'". It feels like you get a lot of protection this way.

Why MCPs will continue to be adopted

So where are they useful? Although everything I've mentioned is true there are some angles I see that give MCPs a clear edge. Firstly, MCPs all adhere to a common format unlike clis. Do you want two clis with similar commands both to be in context at the same time? MCPs have context synergy. Next, they have their own "sandboxing" rules. The LLMs "sandboxing" rules prevent it from operating outside their specified folder. If you tell it to call touch ~/foo-sandbox.js it will surface a permission warning. But what if you wanted a tool that had it's own sandbox? MCP server is the way to go. Lastly, if you are using a lot of endpoints that do expose an MCP and you don't want to go through and define a bunch of custom commands, then yeah use them. Having a marketplace of prebuilt interfaces for your LLM to just install is... kind of amazing.

Overall, I think clis are the way to go for most things, but MCPs do have a place in the broader ecosystem and would bet they continue to be come popular.