🤖 WebMCP: how websites are becoming agent ready

A recent episode of The Startup Ideas Podcast walked through WebMCP, and it is the clearest picture yet of what an agent friendly website looks like. Here is the summary, written for people who build or run websites rather than for protocol nerds.

What WebMCP is

WebMCP is the Model Context Protocol brought inside the browser interface. Microsoft and Google worked on it together, and it landed in Chrome as an experimental feature in February. Your site publishes a short list of tools. Whatever agent the visitor already uses, whether that is Claude, ChatGPT or something else, reads that list and calls the tools it needs.

The older approach was far heavier. An agent would pull the whole page structure, scan the code, take screenshots and hunt for the right button. That is slow, and it breaks the moment you ship a redesign.

The demo that makes it click

The example used was an espresso store. The shopper typed a plain request: upgrade my machine so I can make two flat whites a day, with about 32 centimetres of counter space. The page responded on its own, put two machines side by side, and highlighted the exact specifications behind the request.

Asked which accessories fit one of the machines, the store exposed tools for reading the customer's existing gear, reading the cart and applying coupons. The agent suggested a matching water softener, added it to the cart on request, and applied a ten percent discount, all while the normal shopping interface stayed on screen.

The login is the clever part

API keys, tokens and identity plumbing are what keep most stores from going agent friendly. WebMCP sidesteps that by running on the browser session. The visitor logs in with a normal username and password, and the agent inherits that session as it stands.

That also means the tool list can be conditional. Logged out visitors might see three tools. Logged in visitors see the full set. It is the least technical route to an agent ready store, and it still carries the most features.

Where WebMCP sits on the spectrum

• Raw API: headless, and the agent needs a key.

• MCP server: headless, and another layer to run on your backend.

• Computer use: the agent screenshots and clicks like a person. Slow.

• Browser MCP: the agent reads the page markup and guesses where to click. Fragile.

• WebMCP: the site hands over a tool list, and the visual interface stays.

• In-app agent: your agent, your tokens, your configuration and your bill.

Why this matters over the next few years

The expectation is that over roughly the next 24 to 36 months, visitors turn up with their own agent already loaded with their context and memory, and take it everywhere. WebMCP meets that visitor at the front door without removing the visual interface that ordinary people still want.

Where to build first

• Compatibility driven commerce, where every specification has to match, such as cameras or coffee gear.

• Software admin surfaces: consoles, analytics dashboards and marketing tools.

• Regulated work such as insurance and banking self service, using read-only tools that surface the right information and point people to the right page.

• Internal tools for your own team, which is the easiest place to start today.

The business idea inside it

The obvious service business is making ordinary business websites agent ready: law firms, home services, clinics and similar. Build a first version of the tools that matter, such as request a quote or book a consultation, then charge a setup fee plus a monthly retainer for monitoring and evaluation. The pitch is simple: agents are a new class of visitor, and those visitors book jobs.

The practical takeaway

If you run a website with anything transactional on it, the useful question is no longer whether AI can read your page. It is which five actions you would let a visitor's agent perform directly, and whether your site can hand those over cleanly.

Source: The Startup Ideas Podcast on X, posted 26 August 2026.