My wife and I are moving house, and for the past month we've been flat hunting, which is a whole job in itself when you live in London. So we set our OpenClaw agent Alfred on the task and asked him to provide some options for us to view each day. However, we quickly learned that even though he is a very capable agent and has a robust research and browsing stack, he still struggles with older websites that were not optimised for the age of AI.

He defaults to a mixture of web scraping and browser use, which gets him halfway there (if the website doesn't classify him as a bot), but it's still very much a clunky experience, especially if you'd like to co-browse. At that point, he simply couldn't keep up.

Which led me to ask the question, "If a website doesn't yet have the infrastructure for seamless agent use, how do we give our agents a helping hand in the meantime? How do we provide a dictionary/translator that allows them to ask for the restroom in the middle of Barcelona?" That's quite hard when you don't speak the language.

The Agent Web Gateway was born as a solution. It serves as a stand-in for older, non-WebMCP-compatible websites. Like agent-native subtitles on a foreign language film. The gateway handles a site-specific search, normalises the data, applies the constraints, verifies the results, and returns the best match through that same WebMCP interaction. I've tested it for shopping, rental searches, and job searches with varying success, but the promise is there.

Now, when Alfred tries to browse these websites, he's no longer lost. He doesn't try to brute-force his way through a simple query, burning precious tokens. He speaks to the gateway using WebMCP, it speaks back, he shares his findings with me, and we browse the links together.

My dream is that we will need it less and less over time as WebMCP adoption picks up. But for now, Alfred and I are finally on the same page.

Built With

Share this project:

Updates

Submission history