sophie.tech

What is WebMCP?

S.Allen — Sat, 12 Sep 2026 10:00:00 +1000

I use [[netlify|Netlify]] a lot nowadays, and on their home page I saw they were running a "WebMCP Challenge". I looked up what [[WebMCP]] is, but even after reading about it, it was hard to grasp. So I asked [[claude]] to make an interactive visual so I could understand:

Open the explainer

In short, it's a technology that makes sense. Right now an AI agent has two ways to use a [[website]], and both are clumsy. It can look at the screen and guess, which means screenshots, hunting for the right button, and breaking the moment you redesign anything. Or the site builds and hosts an [[MCP Server|MCP server]], which every user then has to connect to their AI by hand, bloating the context and burning tokens before anyone has done a thing.

WebMCP seems to be a good solution:

The page registers each tool with a name, a description, and the inputs it takes. The agent reads that instead of guessing from the screen.
There is no setup because opening the page is the connection.
The tools run in the page so they use the login you already have.
A page only registers the tools for what is on screen so there is less for the AI to hold in context.
You can see the page changing on your browser while [[AI]] is using the tools.
The browser checks each call, so a site only offers its tools to who it allows.

It doesn't replace MCP. Something with no front end, like a database or an API, still wants a proper MCP server. WebMCP is for the things people normally do by clicking.

It is still early days. It's a draft standard rather than a finished one, Edge ships it today and Chrome has it behind an origin trial, and the API has already changed once this year.

Latest in Local - Image Generation

S.Allen — Tue, 08 Sep 2026 10:00:00 +1000

I have played around with many [[local AI]] tools at this point, starting with using local [[LLM]]'s (such as [[deepseek]], [[qwen]], [[llama]] etc etc), then moving to [[Text-to-speech]], and then the slightly harder [[Speech-to-text]]. For a while I avoided [[image generation]] since I thought it would be immensely hard and need a powerful gpu, but yet again, I was proved that local AI is much easier than expected.

Since I had my [[laptop]] at the time, which is a [[2021 Macbook Pro]] with [[M1]] chip and 16GB RAM, I opted to try the tool Draw Things. The website kept pushing to download from the App store, but I eventually found that the Downloads page had zip files that had the application inside.

After downloading, I consulted [[claude]] to find out which image-gen model to download for my Mac and it suggested [Z Image Turbo 1.0 6-bit]. After putting a prompt, and waiting 5 minutes (which the estimation initially said 30seconds) it came out with an amazing and coherent image! (which soon I will upload)

Keep posted, I have it in the works to make a Hardware Guide and Help Box for localAI website which will help guide and have a chat box to ask what AI and models you can run on your hardware for what task.

IAA Systers Program

S.Allen — Mon, 07 Sep 2026 10:00:00 +1000

Thank you so much for the opportunity to be a part of the [[IAA]][[Systers Program 2026]], what a powerful room of intelligent, hard-working, ambitious women. We will be looking back 10-20 years from now with awe at our journey's and the way we grew.