How MCPs save millions of dollars (vs hosting your own LLM models)
The current trend is to just slap a chat widget on your website that vaguely knows a few things about your product/business then slap “Now with AI” all over the product.
Canva did this and exactly this and created an awful user experience that I am sure damaged their relationship with the customer.
The chat widget was basically an expensive and slow way to search their templates, beyond that it couldn’t do anything.
I say expensive, not because I personally had to pay a penny more but because every interaction with their chat widget they had to spin up some LLM model to interpret the request and respond. And unless they are using the tiniest of models then that bill ramps up pretty quick. They might have been using a tiny model though, it really was not great.
What is my solution:
Stop slapping chat widgets on everything. Instead optimize your site so ChatGPT, Claude, Gemini can interact with your website. Let people bring their own agents/models which they pay for, not you. Then just let the users' Agents select the Canva templates and edit them.
The user’s Agent has all types of context that your chatbot could never have. Depending on which agentic software you chose it has access to every conversation the user has ever had with the agent. It likely has access to their Google drive or your filesystem of choice which has all the previous marketing material to draw from and pipe into Canva.
The user experience will be a lot better and you won’t have to pay the bills on spinning up expensive LLM Models.
This all sounds great in theory Matt but how do I do this?
Why with MCPs and Tool Calls of course.
I know I sound like a broken record with this MCP/Tool call stuff right now but being as it is my job to figure out how to scale up my clients infrastructure to service millions of users needs with out breaking the bank I really wanted to point out how expensive and not scalable slapping a chat widget on your product could be vs just exposing the tools for their agents to use it.
I am not saying there is never a good use case for a chat widget but it shouldn’t be the default. Besides, if you build out the MCPs / Tools first and make that functional then if you decide down the line to add a chat bot it will have access to your extensive library of tools which will give it more functionality to wow your customers.
I guess what I am trying to say is take a “Tool first” approach to AI. That is it! I am coining the term “Tool first”. You heard it here first… I hope.
If you are interested in learning more about this, join my workshop this Friday.