ChatGPT browser adds WebMCP support for sites as tools

• Compatible websites can now become callable tool surfaces for ChatGPT and Codex in the desktop app’s built-in browser and ChatGPT Sites, according to OpenAIDevs' rollout.
• WebMCP gives sites structured actions instead of leaving agents to infer a UI, as OpenAIDevs' definition describes the experimental standard.
• OpenAI’s early reference experience is a browser-native 3D workspace, identified in pvncher's demo note as Codex Modeling Studio.
• The 10-day challenge pairs OpenAI’s prize pool with sponsor incentives, including the credits in vercel_dev's offer.

The WebMCP specification gives pages both declarative form annotations and JavaScript-defined tools. Chrome’s security guide already specifies untrusted-content hints, cross-origin controls, and output budgets. Codex Modeling Studio is the first demo OpenAI has put in front of builders.

OpenAI calls its implementation “site tools.” Its ChatGPT Site Tools announcement says a site can expose functionality and data alongside the ordinary interface, while the human and agent share the live page and signed-in session.

OpenAI names the desktop app’s built-in browser and ChatGPT Sites as the launch surfaces. On a compatible site, ChatGPT Work and Codex can discover the tools available on that page.

pvncher’s WebMCP explanation frames the interaction as a collaborative one, with the person and Codex working against the same content in the built-in browser. Codex Modeling Studio applies that pattern to authoring 3D scenes.

Chrome documents WebMCP as a proposed browser API, not a remote connector. Its overview offers two ways to turn a page capability into an agent-callable contract:

1. Forms: human-interpreted controls → annotated toolname and tooldescription plus fields as parameters, declarative tools, per Chrome’s declarative API guide.
2. Application workflows: UI traversal → document.modelContext.registerTool() with a name, description, and input schema, imperative tools, per the imperative API documentation.
3. Tool discovery: page semantics inferred at runtime → document.modelContext.getTools(), a browser-supplied inventory, in the same imperative API.

The imperative path is intended for operations including form input, navigation, and state management. OpenAIDevs’ flight-search demo is the product version of the pitch: an agent fills a travel form through declared site actions.

OpenAI and Chrome describe the performance target, but neither the challenge page nor Chrome’s WebMCP overview publishes a task-success score.

First-party

• OpenAI WebMCP task-success rates reported: 0 → 0, 0-point reported delta, across the announcement and challenge materials.

Third-party evaluators

• Independent WebMCP task-success rates reported: 0 → 0, 0-point reported delta, in Chrome’s linked implementation documentation.

Customer-reported

• Customer WebMCP task-success rates reported: 0 → 0, 0-point reported delta, in the supplied launch evidence.

OpenAIDevs’ flight demo is a worked task, not a benchmark suite. The challenge resource list instead directs builders to test deployed apps in ChatGPT’s in-app browser or in Chrome with the WebMCP testing flag enabled.

The standard is a Community Group work item rather than a W3C standard, and Chrome labels its implementation an origin trial. With no published baseline, the rollout provides no numerical regression comparison.

Chrome’s security material treats the threat model as active: indirect prompt injection remains possible because models consume instructions and untrusted data in one token stream. Its guidance includes:

• untrustedContentHint for tools returning user-generated or external content.
• readOnlyHint for tools that do not change state.
• Default same-origin isolation, with exposedTo available only for named trusted origins.
• Recommended ceilings of 500 characters for a tool description and 1.5K for an individual output.

Existing ChatGPT browser reports show separate rough edges around access and sessions. In koltregaskes' Android report, the cloud browser could reach form fields but not bring up a keyboard, while nicdunz' sign-in test reported inconsistent behavior in ChatGPT Work. Neither post demonstrates a WebMCP-specific failure.

OpenAI opened registration and submissions on August 25 at noon PT. Its challenge page lists September 3 at 1 p.m. PT as the deadline and September 23 for winners, subject to change with submission volume.

There is a deadline discrepancy: OpenAIDevs' submission post says September 3 at 5 p.m. PT, while the official page gives 1 p.m. PT. The official prize list gives each of the top 10 projects $3,000 cash, a year of ChatGPT Pro, a Codex Micro keyboard, swag, and partner prizes.

• The OpenAIDevs kickoff post advertised the opening livestream for product demos, examples, prizes, and entry details.
• Vercel says the first 1,000 builders get $20 in AI Gateway credits, while its top 10 receive $3,600 in Vercel credits and $600 in AI Gateway credits in vercel_dev's announcement.