I am loving Model Context Protocol but there is one massive issue I have with the way I am currently seeing it being implemented: The ability to discover new services and add them to your LLM Agent’s repertoire of data sources/commands is GARBAGE…. At least for now but keep in mind it is in Beta.

Right now for Claude Desktop I have to create a JSON file that is buried deep in some Apps directory. Then I have to manually add each service. Potential users, like my mother, are going to struggle to adapt to this.

Obviously this is not the end goal so lets dream for a minute and see if we can see what a better process might look like.

Let’s say you already shop at the Acme Warehouse Store website quite frequently. You tell your LLM agent that and it stores your preferences away in its memory. Then, because there is an internet standard similar to robots.txt it checks mcp.txt which tells it the URL that it can connect with to get access to the MCP server run by Acme Warehouse Store.

Now, without any real technical knowledge on behalf of the human, the LLM Agent has a structured way to communicate with and query everything Acme Warehouse Store has to offer.

If you are concerned about security you are 100% right. What happens if the LLM goes rogue and starts to make purchases at random? How does it have access to my Acme Warehouse Store account? But I will tackle that in another post because MCP has sweeping implications on how that could affect all things security wise. Lets move on.

So in this scenario you can give it a website you like and the LLM agent can automatically connect to it. Great but how do you know what websites support MCP? Aggregators are already popping up for this.

Once the MCP discovery is implemented you better bet that MCP aggregators will be popping up next. And by that I mean an MCP server that the LLM agent can query to find other MCP servers. It will be like the Google of MCP servers but without the need for a text box or any UI at all. Your LLM Agent of choice will be the interface.

If all this doesn’t already exist it probably will by the time I finish typing this sentence so if you know of an implementation, protocol, or standard like this please let me know about it. If you want to chat with me about it please sign up for my Workshop on how to run AI/ML on AWS this Friday at 1PM Central and we can chat about it then,