Key details

  1. OpenAI launched the WebMCP Challenge on August 25, 2026 with Google Chrome, Cloudflare, Shopify, Vercel, Render and Netlify among the supporters.
  2. OpenAI describes site tools as ChatGPT’s implementation of the proposed WebMCP standard.
  3. ChatGPT’s desktop built-in browser can discover WebMCP tools when both the account/model and current page support them.
  4. Chrome exposes WebMCP through an origin trial and experimental flag, not as a stable universally shipped web API.
  5. The imperative API uses `document.modelContext.registerTool()` with structured descriptions and JSON schemas.
  6. Chrome deprecated the earlier `navigator.modelContext` surface in Chrome 150.
  7. Cross-origin tool exposure requires both Permissions Policy delegation and explicit origin controls.
  8. WebMCP is designed primarily for local browser workflows with a human in the loop; headless use is not its primary design target.

What builders should take away

  1. If your web app has high-value workflows that browser agents currently click through unreliably, prototype one narrow WebMCP tool around an existing server-authorized action rather than exposing the whole application at once.
  2. Treat tool descriptions and schemas as an agent-facing API contract: validate every argument server-side and never use the browser tool definition as the authorization boundary.
  3. Require explicit user confirmation for purchases, sends, deletes, account changes or other consequential mutations even when the agent selected the correct tool.
  4. Keep ordinary UI and API paths working. WebMCP is a progressive enhancement and remains an experimental browser standard.
  5. Test with both ChatGPT’s supported built-in browser flow and Chrome’s origin-trial implementation if cross-client behavior matters; do not assume experimental implementations are identical.
  6. Audit embedded and cross-origin components carefully because tool delegation introduces another capability boundary alongside cookies, authentication and normal iframe permissions.

What changed

On August 25, OpenAI made WebMCP materially more accessible to builders by launching its WebMCP Challenge and documenting site tools in ChatGPT’s desktop built-in browser. WebMCP is a proposed open web standard already available through a Chrome origin trial: a page can register named tools with descriptions and JSON input schemas, and a compatible browser exposes those tools to the user’s agent. OpenAI says ChatGPT site tools implement WebMCP and can be discovered by ChatGPT in its built-in browser; Chrome exposes the same proposed standard through its experimental implementation. The result is a shared path for agent-native web actions that does not require the agent to infer every step from screenshots or DOM clicks.

Why it matters

Browser agents are usually forced to operate a website as if they were a human with a mouse and keyboard. That makes automation slow and fragile when layouts change, elements are ambiguous or multi-step state is hard to infer. WebMCP lets the website declare what actions are available and execute them inside the page’s own application logic and signed-in session. For SaaS and web-product builders, that creates a new integration surface: instead of maintaining a separate external API solely for agents, they can progressively expose selected user-facing capabilities directly from the live web application. The trade-off is that these tools can perform real actions with real user state, so permissions, confirmation, untrusted content and tool design become part of browser security rather than merely UX.

The website defines the contract

WebMCP’s imperative API lets a page register tools through `document.modelContext.registerTool()`. Each tool has a name, description, JSON input schema and an execution handler. Chrome also documents a declarative path for suitable form interactions. A compatible agent sees the structured contract rather than reverse-engineering what a button or field probably does.

ChatGPT now provides a real client path

OpenAI documents site tools in the ChatGPT desktop app’s built-in browser. When a supported page exposes a matching WebMCP tool, ChatGPT can discover it while the page is open, operate within the same live signed-in session and ask for website-access permission before continuing. OpenAI also says ChatGPT Work and Codex can use WebMCP-enabled site tools in supported browser workflows.

Chrome is still treating WebMCP as experimental

Chrome’s implementation remains an origin trial and local-development flag rather than a stable cross-browser API. The current interface uses `document.modelContext`; the earlier `navigator.modelContext` form was already deprecated in Chrome 150, showing that the API can still move. Chrome’s August documentation also adds lifecycle, cancellation and cross-origin controls, including Permissions Policy and explicit origin exposure for tools in embedded contexts.

This is different from a remote MCP server

WebMCP borrows the tool-oriented idea familiar from Model Context Protocol but places the capability inside the current web page and browser session. That is distinct from exposing a standalone remote MCP server. A site can therefore use its existing frontend state, authentication and application code, while the browser mediates discovery and invocation for the user’s agent.

Tool safety becomes application security

A WebMCP action may search, edit, submit, purchase or mutate state, depending on what the website exposes. Builders need narrow schemas, clear descriptions, appropriate read-only or untrusted-content annotations, server-side authorization and human confirmation for consequential actions. Chrome also restricts WebMCP to origin-isolated documents and gates cross-origin use through the `tools` Permissions Policy.

What to watch next

  • Whether WebMCP progresses beyond Chrome’s origin trial toward a stable multi-browser standard.
  • Whether Safari, Firefox or other browser engines signal implementation interest.
  • How OpenAI expands site-tools availability beyond the current built-in-browser/account constraints.
  • Whether common frameworks and SaaS platforms add first-class WebMCP abstractions and security conventions.
  • Evidence that structured tools materially improve task-completion reliability versus browser actuation on real production workflows.

Still unclear

  • WebMCP is explicitly experimental and its JavaScript surface has already changed during the trial.
  • OpenAI site-tools availability depends on account, model and browser support and is not equivalent to universal ChatGPT web support.
  • There is not yet broad independent production evidence quantifying reliability gains or new security failure modes at scale.
  • Cross-browser adoption is not established.

Sources

Direct reading behind this dossier.

4 sources
WebMCP Challenge
OpenAI primary announcement

OpenAI launch surface establishing WebMCP as an experimental open standard, ChatGPT built-in-browser testing support, challenge dates and participating ecosystem vendors.

Using site tools in the ChatGPT desktop app
OpenAI Help Center primary documentation

Documents ChatGPT site-tool discovery, same-page browser use, permission prompts, availability constraints and WebMCP relationship.

WebMCP
Chrome for Developers primary documentation

Chrome overview of the proposed standard, origin trial, imperative/declarative APIs, use cases, limitations and security model.

WebMCP Imperative API
Chrome for Developers primary technical documentation

Current tool-registration API, document.modelContext migration, cancellation and cross-origin controls; updated August 20.