OpenAI has added support for WebMCP, a draft standard that lets a webpage register JavaScript functions as callable tools for an AI agent, to the built-in browser in its ChatGPT desktop app. A site that implements it gets a new lever: it can define exactly which actions ChatGPT is allowed to take on a page, rather than leaving the agent to infer them by clicking and typing. Search Engine Journal reported the rollout on Thursday, citing OpenAI’s own documentation.
This is not the server-based Model Context Protocol that ChatGPT has supported since 2025, and it is not a robots.txt directive or a crawling permission. WebMCP tools live inside a signed-in browser session and vanish the moment the page closes; they never touch what a crawler indexes or what a search bot is permitted to fetch. That distinction matters for search teams who might otherwise lump every mention of “MCP” into one bucket. Robots.txt governs whether a crawler may request a URL at all. WebMCP governs what a live agent, already on the page inside a human’s session, is allowed to do once it gets there.
Registered tools handle page-level actions: pulling up a document, opening a file for editing, navigating a dashboard, weighing options on a travel search, or adding an item to a cart. The exact menu of tools changes from one site to the next, and a tool used on one page will not carry over to another. ChatGPT’s browser marks availability with an icon in the address bar, distinguishing tools that only read a page from tools that can change it.
Access is deliberately narrow right now. The capability currently needs the GPT-5.6 Sol or Terra model; the Luna model has it turned off entirely, and the feature does not appear at all inside Enterprise or Edu accounts. It also does not function in Chrome when reached through ChatGPT. Developers who want to test it in Chrome directly need an experimental flag or a spot in Chrome’s origin trial, which puts practical reach well behind the announcement itself.
The underlying specification is still a draft from the W3C Web Machine Learning Community Group and has not entered the W3C Standards Track. OpenAI is building a shipping product on top of a proposal that has not been ratified, so the tool schema and behavior a site implements today could shift before the standard settles. Search Engine Journal noted that Google’s developer guidance already pushes site owners toward agent-friendly design; WebMCP gives that guidance a concrete, code-level target instead of a general principle to satisfy.
What remains unstated is the visibility payoff. Per Search Engine Journal’s review of OpenAI’s documentation, nothing indicates whether registering site tools changes a page’s odds of being cited, ranked, or surfaced inside a ChatGPT answer. The publication also flagged a standing security gap: Chrome’s own guidance already treats malicious tool descriptions and manipulated tool outputs as prompt-injection vectors for browser agents, and OpenAI’s safety review runs per tool call rather than certifying a site as trustworthy in advance.
For search and product teams, the near-term move is measurement, not adoption. WebMCP opens a new, site-controlled channel for agent interaction that sits outside both crawling and robots.txt, so any team running e-commerce, travel, or document-heavy pages should track whether exposing tools changes ChatGPT’s behavior on their site before committing engineering time to a specification that is still a proposal.
Search Engine Journal (Matt G. Southern) reported the WebMCP rollout on August 27, 2026.
