It's time to critique MCP
Recently in the MCP Cohort Discord channel, a learner shared the blog post “Code execution with MCP: Building more efficient agents” by Anthropic which presents similar findings to Cloudflare’s “Code Mode: the better way to use MCP”.
He also linked to a video that used Anthropic’s article to take shots at MCP and wanted to know my thoughts.
My first thought: we’ve definitely entered the critique phase of MCP!
There’s a difference between the social media chatter that loves to declare the death of one thing for another thing to become useful, let alone popular, but I don’t believe we are playing a zero sum game like that.
So let’s explore the reality of the critiques, and what we can and should learn from them.
Understanding the Critique
To quickly summarize the articles linked above, engineers at both Anthropic and Cloudflare have identified some significant inefficiencies in implementations of MCP Clients leading to a poor user experience as a result of context bloat, security and privacy issues, state management, and performance.
Thing is… none of these issues are a surprise to me, or many of the people working closely with these tools!
In fact, I noticed all of the same things when I started getting into MCP back in March (around the time that Authentication was added to the spec).
My take: the need for these kinds of issues and necessary optimizations were obvious from the start, not a recent discovery.
And most importantly, I think it’s GOOD that more people are talking about these issues. That’s how we create better long term solutions and amazing user experiences!
The real question is, what have we learned?
I think that the future of user interaction will continue to move more user experiences from standard browsers and websites into agents and MCP-backed services.
Because of this shift (that’s already happening), I often make parallels between early days of the web.
When web browsers got started, there were TONs of problems with them and the websites they accessed. They were insecure. They were slow. They were inefficient. Sound familiar?
Over time, standards and implementations evolved, but browsers never stopped solving problems. That’s why they are the dominant player in the world for connecting users to our services.
Now, we have a new range of experiences to create, and we get to go through the whole cycle again. But with the knowledge we earned from last time, we can do it better and faster!
Using the Epic Principles to Guide Us
So when I see articles like the ones from Anthropic and Cloudflare, and the broader discourse comparing MCP to other techniques that are being built in parallel, I see them ALL as evolutions of the platform I believe we’ll all be building upon in the near future (many are already building on it actively).
And I bring it back to one of my favorite of the Epic Principles: Make it work, make it right, make it fast.
First make certain that what you're trying to do is possible (make it work), then make sure what you're trying to do accomplishes the objective correctly and handles edge cases (make it right), and then figure out a way to make that fast.
MCP definitely works (I and many others use it extensively every day!) so I’d say that we’re somewhere between the “make it right” and “make it fast” stages of this idea.
Another thing to consider is that the vast majority of developers building on the web are working on websites rather than web browsers. It’s the same thing with MCP.
The vast majority of you will be working exposing your service to AI Agents via MCP Servers, not on building AI Agents that consume your services via MCP Clients.
So for now, my focus is primarily on the MCP Server aspect of this platform. That’s where most of us will work.
And when I DO see innovations happening on the client side, it’s just further validation to me that the UX of people using my MCP server is only going to improve over time.
We’ve already seen that play out in MCP over the last year (that’s right, we’re coming up on MCP’s birthday on November 25th!). The spec has continued to evolve. The right players are actively participating in both the spec and implementation of clients and servers.
Recent optimizations like Cloudflare’s “Code Mode” (or as Anthropic calls it, “Code Execution”), but this is not an admission that MCP is failing.
On the contrary, the MCP standard makes these kinds of optimizations possible because it gives a standard interface for services to define tools, and then optimizations like code mode can build on that to make the LLM write code that invokes those tools as efficiently as possible.
To learn more details on how this works, check out the articles above. It’s awesome to see.