How OpenAI’s WebMCP Lets Websites Expose Tools Directly to AI Agents

Introduction

In August 2026 OpenAI released the Web Model Context Protocol (WebMCP) together with a Chrome side‑panel extension that embeds a Codex agent. WebMCP lets a website publish a well‑known JSON manifest so that AI agents in the browser can discover and invoke the site’s capabilities automatically.

What WebMCP Is

WebMCP differs from Anthropic’s MCP (released late 2024) by moving tool registration from the AI developer to the website operator. A site places a /.well-known/webmcp.json file that enumerates tools such as “fetch full paper”, “extract citations”, or “export BibTeX”. The browser‑based agent can then call those tools directly.

Technical Architecture and Workflow

The protocol follows four steps: discovery, negotiation, invocation, and response.

Discovery – the agent fetches /.well-known/webmcp.json via the standard .well-known path.

Negotiation – the manifest’s JSON Schema describes each tool’s name, description, endpoint, HTTP method, parameters, and authentication.

Invocation – the agent sends a standard HTTPS REST request to the declared endpoint.

Response – the service returns JSON data to the agent.

A minimal manifest example:

{
  "schema_version": "1.0",
  "provider": "example.com",
  "tools": [
    {
      "name": "search_papers",
      "description": "Search papers by keyword and return a list of abstracts",
      "endpoint": "/api/webmcp/search",
      "method": "POST",
      "parameters": {
        "query": {"type": "string", "required": true},
        "limit": {"type": "integer", "default": 10}
      }
    }
  ],
  "auth": {"type": "oauth2", "authorization_url": "/oauth/authorize"}
}

The schema, combined with standard HTTPS, OAuth2 or API‑key authentication, and CORS headers, enables cross‑site tool aggregation.

Core Differences from Anthropic MCP

Runtime location : Anthropic MCP runs on a local or remote server process; WebMCP relies on a website‑side declaration discovered by the browser.

Discovery mechanism : MCP requires manual server address configuration; WebMCP uses automatic .well-known discovery.

Transport protocol : MCP uses JSON‑RPC over stdio/SSE/streamable HTTP; WebMCP uses standard HTTPS REST.

Tool registration party : MCP tools are registered by AI application developers; WebMCP tools are registered by website operators.

Authentication model : MCP handles auth server‑side; WebMCP standardizes OAuth2 or API‑Key handling.

Ecosystem positioning : MCP is a generic tool‑connection layer; WebMCP is a web‑native tool exposure layer.

Chrome Extension Practical Test

The side‑panel appears on the right side of Chrome, detects whether the current page provides a WebMCP manifest, and lists available tools for natural‑language invocation. Measured response latency is typically 1–3 seconds, limited by network round‑trip and the target site’s API speed. GPT‑4.1 drives tool selection and parameter filling.

Four In‑Depth Scenarios

Paper Analysis

On an arXiv page (which currently lacks a WebMCP manifest), the extension falls back to page‑content extraction plus long‑text analysis. For a 12‑page Transformer‑variant paper the agent produced an ~800‑word structured summary covering problem definition, method highlights, and experimental conclusions, with no obvious hallucinations. If the site exposed WebMCP tools such as “fetch full text” or “extract figures”, the agent could retrieve structured data directly.

Image Understanding

On Unsplash, the side‑panel invoked GPT‑4o’s vision capability. It described a complex scene (“a climber in a red jacket at dusk on a cliff with layered clouds”) and identified composition techniques (“rule of thirds, subject at right‑hand golden ratio”).

Webpage Translation

Translating a Japanese technical blog to Chinese yielded higher contextual fidelity than Google Translate. For example, the term 「デプロイ」was rendered as “publish” when the surrounding context implied a release rather than a generic deployment. Translating a ~3000‑word article took about 8 seconds.

Notion Integration

Notion adopted WebMCP in August 2026. The side‑panel automatically displayed tools such as “create page”, “query database”, and “update property”. An instruction to “organize the current page’s to‑do items into a table with priority” resulted in a new table‑view database populated with extracted items and priority fields, using Notion’s existing OAuth login without requiring additional API tokens.

Developer Integration Guide

Create /.well-known/webmcp.json at the site root following the schema.

Implement each declared tool as a standard REST endpoint that accepts and returns JSON.

Configure authentication, preferably OAuth2, so the browser can reuse the user’s login session.

Add appropriate CORS headers to allow requests from the Chrome extension’s origin.

A seasoned backend developer can complete these steps in half a day to one day. OpenAI provides a webmcp-validator utility to verify manifest compliance. The protocol also defines an optional rate_limit field that lets sites cap call frequency per tool.

Open Issues

WebMCP is still early; few sites have adopted it. Open concerns include security (malicious sites exposing harmful tools), privacy (data flow when agents invoke APIs on behalf of users), and ecosystem fragmentation (potential competing standards from Google, Anthropic, etc.).