- An OpenAI presenter said WebMCP lets an agent use a site through exposed tools, where computer use has to click through the screen.
- The presenter urged builders to make apps work well with agents, though the talk gave no speed or cost figures.
A presenter at OpenAI's DevDay 2026 showed an agent using a site through tools the site exposes. The presenter said this can be faster and cheaper than clicking, depending on the use case. The speakers were not named in the transcript. The point: a site can publish tools for agents next to the pages it shows to people.
What was said
The presenter called WebMCP a new browser standard. Any app can expose tools through it. They work in a WebMCP-enabled browser, such as ChatGPT's in-app browser.
The comparison was with computer use, where an agent operates a site like a person, opening each catalogue item to read its details. A speaker on the show put the difference this way: "the agent can use a site via tools instead of clicking through the UI." With WebMCP, the agent can query the backend directly.
The demo was a gift box for a colleague. The agent searched Slack to learn what the colleague likes. It then used the site's tools to put catalogue suggestions into the live page. The presenter removed a few items to stay under a spending limit, added the rest to the basket and confirmed by hand. The talk did not say whether tools could have done those steps. The agent also drafted a personal message and put it in the page. The presenter urged builders to make apps that work well with agents.
Why it matters
Our reading: a site now has two front doors. A page is made for a person to read and click. A tool list names actions an agent can call without clicking.
The demo did not settle which actions belong behind tools and which stay with the person. That choice sits with the site.
The presenter said the agent works with the user's own data and apps when it calls the tools. Each tool is therefore a decision about reach. A budget signer can ask: which actions do we list for agents? What user data can an agent reach through them? Which steps still need a person to confirm?
The other side
This is a product demonstration from the company that sells the tooling. The speed and cost advantage came with the hedge "depending on the use case", and no measurements were given. The demo ran in ChatGPT's in-app browser. The talk did not say which other browsers support WebMCP.
The excerpts also left open how a site decides which actions are safe to expose, and how it tells a trusted agent from a harmful one. The presenter said no extra work is needed, but that applied to sites built with this product.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- OpenAI: Sites and the Next Generation of Web Apps | DevDay 2026 (2026-11-11)
