Why I Even Looked at webMCP
This post started with a conversation with Diana Dreher from Microsoft.
At first, I honestly did not get the point of webMCP. I had the usual thought: “Why do I need another protocol if agents can already click buttons like a caffeinated intern?”
Then we had a longer discussion about where agents are still weak today:
- They are often forced into fragile RPA-style browser automation
- They struggle with reliable intent-to-action mapping on complex UIs
- They have no consistent contract for “what can I do on this website?”
We also talked about the business side and monetization opportunities, but this post sticks to the technical side.
The Experiment: Build a Real Sample
I decided to test the idea with a practical app instead of theorizing forever.
I asked Claude Code to generate a sample Contoso-style shop and then enhanced it with:
- Better logging, so I could actually see what was happening
- A webMCP implementation on top of existing capabilities
Sample repo:
After about one to two hours, I had a working shop and could start testing the agent flow end to end.
What Actually Worked
And yes, it is kind of cool.
Through the Claude browser extension, I could:
- Search products
- Add products to the cart
- Update user profile data
- Execute checkout
One important detail: I had to explicitly tell Claude to check for a webMCP implementation first. If I did not, it often tried the default UI-driving path.
I also tested it with the Claude desktop app. It still had to spin up a browser session, but the flow worked.
The Part That Blew My Mind
Making the shop “agentic” was much easier than expected.
Most sites today bolt on a chat assistant in the bottom-right corner with a custom UI and awkward handoff patterns. You know the one.
With webMCP, I could skip that pattern entirely for agent clients:
- Open the chatbot panel in the browser
- Tell it to use webMCP
- Let the agent interact through explicit capabilities instead of guessing UI behavior
Right now I only have this running nicely with Claude, not Copilot, but the direction is very promising.
Why This Matters for Existing Websites
The most practical takeaway: I did not rebuild core shop logic.
I told the coding agent to reuse what already existed and expose another interface for it. That means a lot of current websites could be adapted without a full rewrite.
In other words:
- Keep existing business logic
- Add a machine-usable contract layer
- Gain a cleaner path for agent interactions
Conclusion
Fun project, solid learning.
If webMCP discovery becomes more automatic in clients, this gets even better.
Given how quickly this worked in a real sample, I now see webMCP as a very practical way to make existing web apps more agent-friendly without adding another awkward chatbot bubble to every page.