WebMCP gives AI agents a way to use your website
Published September 16, 2026
Learn how WebMCP lets websites expose tools to AI agents, with a working nine-tool demo, real-world ecommerce use cases, and insights into GEO, browser support, and implementation.
I spent about an hour adding nine WebMCP tools to my personal website. Then I asked an AI agent questions about my work. It could find specific projects and return links that supported its answers.
That small experiment made me think about ecommerce. A customer could ask a question, and an agent could use tools supplied by the retailer. The retailer could connect those tools to the services its website already uses.
WebMCP is an experimental browser API that lets websites offer tools to compatible AI agents. A tool describes an operation and the inputs it accepts. The website supplies the code that carries out the operation.
For example, "Has Peter used Ollama?" returned my AI Chatbot project. A question about Next.js returned Live Session Lab and HHPC2. These answers connected questions to evidence from my published project catalog.
A practical starting point for ecommerce
My site uses tools such as search_projects and get_project. The same pattern could support product discovery. A retailer could define a focused tool that accepts a customer's filters and returns matching products.
The following examples are possible applications, not features I built in this experiment.
"Which running shoes cost under $150?"
A retailer could provide a product search tool with a category and a maximum price.
"Is this available in size 10?"
A retailer could provide a tool that checks the selected variant against current stock data.
"Can I collect this near my ZIP code?"
A retailer could provide a tool that checks store availability for that product and location.
"How do these two products differ?"
A retailer could provide a tool that returns comparable specifications and product links.
These tools could call existing website services when those services already support the operation. The customer and agent could work within the same browser session. The WebMCP explainer describes this use of existing application logic and shared page context.
The existing website does much of the work
A separate MCP server is not required for this approach. A JavaScript tool can call an existing search endpoint and return the result. That can keep the site interface and the agent connected to the same product data.
The amount of work depends on the website. Tool registration is only one part. The operation still needs input checks and useful error handling. Existing access rules must continue to apply. A purchase also needs a clear confirmation step.
In my JavaScript implementation, most tools read public data already embedded in the page. One tool fetches live GitHub repository metadata. The project catalog contains reviewed snapshots, so its results are not live scans of every repository.
That distinction matters in ecommerce. A product description and an inventory check have different freshness requirements. The tool should make the source and limits of its answer clear.
Where GEO and WebMCP fit
My shorthand is: GEO helps people find what a site says. WebMCP lets an agent use what a site does.
GEO means generative engine optimization.
WebMCP tools become available when a compatible browser or agent visits the site. Publishing tools does not guarantee discovery in AI search. I see WebMCP as another way to use a website after reaching it.
Where you can use WebMCP today
Support checked on September 15, 2026. I could not get WebMCP to work with GPT-5.6 in my test. It worked with GPT-6 Astra in the ChatGPT desktop app, using its built-in browser. The screenshot records that working setup.
OpenAI's Site tools documentation also lists GPT-5.6 Sol and GPT-5.6 Terra. It identifies the built-in browser as the supported location for ChatGPT Work and Codex. It excludes GPT-5.6 Luna and Enterprise or Edu workspaces. Access also depends on rollout and the current page.
That page does not yet list GPT-6 Astra. My experience therefore adds a specific observed result to the published guidance. Readers should check the current documentation for their own setup.
OpenAI currently supports JavaScript tools registered in the top-level page. Its documentation excludes tools defined through HTML form attributes and tools inside iframes. Browser support alone does not mean every agent supports every WebMCP feature.
Chrome's documentation provides another testing route: an origin trial from Chrome 149 and a local development flag. Its Model Context Tool Inspector extension can inspect and call tools. The project's implementation status also lists an Edge 150 origin trial and experimental Brave Leo support.
A working experiment with an evolving specification
WebMCP is broader than one AI product. However, ordinary MCP support does not automatically give an agent access to browser tools. The browser and agent need a compatible integration.
The WebMCP specification remains a Community Group draft. It is not a W3C Standard. I would treat a working demonstration as a reason to test a focused use case. It does not establish a date for broad adoption.
For an ecommerce team, I would start with a read-only product search. Check whether the agent selects the right tool and returns accurate results. My nine-tool experiment showed me enough to make that first test worthwhile.
Follow progress and readiness
- Chrome developer guide - Setup, testing tools, and links to the origin trial
- Chrome Platform Status - Chrome implementation and release milestones
- Browser and agent implementation status - Support across products
- WebMCP specification and issues - Proposed API changes and open discussions
- OpenAI Site tools - Supported environments and current limitations
- My webmcp.js - The code behind this experiment