This weekend I was working on a live demo for the presentation I am doing next week.

Since my talk is on Tool Calls, MCPs and how they will be changing how we interact in the future with the internet I wanted to see if ChatGPT, the agentic interface had made any strides towards integrating MCP.

Additionally ChatGPT pro was offering 1 month for $1. I decided to give it a try. I got it downloaded and tried to point it at my own MCP server running locally.

To my surprise and dismay it appeared not to be able to communicate with any local servers. It didn’t even support STDIO.

I pushed my code up to my dev env and tried pointing it at that. It could communicate but kept erroring out saying that it was an invalid format and linking me to their docs.

As it turns out they don’t really support MCP and 3rd party tool calls. They only support 2 methods:

search which takes only 1 text query parameter and fetch which allows the agent to query details about something that was searched.

What a waste of an implementation. I am not sure why you would half-ass such a powerful tool.

Perhaps the lawyers got involved and they wanted to add some safety rails to ensure the users don’t connect to a malicious MCP server but being as you had to be fairly technical to add them in the first place I don’t think that is the case.

The tinfoil hat theory is that they are limiting 3rd party tools capabilities so they can either build their own or give preferential treatment to their partners.

Either way I doubt I will stick with the proplan after this month. The good news is that there is more room in the market for aspiring devs to build their own Agentic software that will fully implement MPC spec instead of limiting it to just 2 methods.

If I wasn’t so busy with my 10 other projects I would be tempted to build one of my own.