MCP, A2A, ACP and UCP explained for business owners
What an AI agent is
An AI assistant answers questions. An AI agent goes one step further: it does something for the person. It checks opening hours, looks at a booking calendar, fills in an enquiry form or puts a product in a cart.
ChatGPT, Gemini, Claude and Copilot can all work this way for some tasks, and web browsers are starting to do the same.
To do a task, an agent needs a way in: agreed rules for asking a system what it can do, and for doing it safely. Those rules are called protocols. Four names come up again and again.
The four protocols in plain words
Who made each, and what it connects
- MCPModel Context Protocol. Anthropic, November 2024.AI toolA serviceThe most used of the four.
- A2AAgent2Agent. Google, April 2025.An agentAnother company's agentVersion 1.0 since March 2026.
- ACPAgentic Commerce Protocol. OpenAI and Stripe, September 2025.An assistantA shop's checkoutStill labelled beta.
- UCPUniversal Commerce Protocol. Google, with Shopify, Etsy, Wayfair, Target and Walmart, January 2026.An agentA shopThe whole shopping trip, from catalogue to order updates.
MCP: how AI tools plug into a service
MCP stands for Model Context Protocol. Anthropic, the maker of Claude, published it in November 2024 as "a new standard for connecting AI assistants to the systems where data lives".
A service runs an MCP server. An AI tool connects to it and sees a list of things it may do, such as "search products" or "book an appointment".
Who runs it now: in December 2025 Anthropic gave MCP to the Agentic AI Foundation, a new home under the Linux Foundation co-founded by Anthropic, Block and OpenAI. The specification has dated versions; the latest is 2026-07-28.
How mature: the most used of the four. At the handover the project counted 10,000 active servers and support in ChatGPT, Claude, Gemini, Copilot and others. Most of that use is in software and developer tools, not on local businesses' websites.
For a small business: you rarely build one yourself. Wix, for example, gives every site a built-in MCP server, so an agent can read the business details, see the services, book or start a purchase.
A2A: how agents talk to other agents
A2A stands for Agent2Agent. Google launched it in April 2025 with more than 50 partners. It lets one company's agent hand a task to another company's agent, for example a travel agent asking a hotel's agent about rooms. Google describes it as a partner to MCP: MCP gives an agent tools, A2A lets agents talk to each other.
Who runs it now: the Linux Foundation took it over in June 2025, and since August 2026 it sits in the same Agentic AI Foundation as MCP.
How mature: the first stable version, 1.0, shipped in March 2026, and more than 150 organisations back it. It is used mostly inside large companies' systems.
For a small business: almost none run their own agent, so A2A rarely applies yet.
ACP: how an assistant checks out with a shop
ACP stands for Agentic Commerce Protocol. OpenAI and Stripe released it on 29 September 2025. It lets an agent start a checkout with a business while the business stays the seller. It is open source and still labelled beta; the latest version is dated 2026-04-17.
Watch out
A reality check: ChatGPT's own "buy inside the chat" feature, built on ACP, was scaled back in March 2026. Today shoppers who find a product in ChatGPT mostly finish on the shop's own checkout. The protocol itself carries on.
For a small business: it matters only if you sell products online, and it reaches you through your shop platform or payment provider. Nobody should hand-write it.
UCP: Google's open standard for shopping with agents
UCP stands for Universal Commerce Protocol. Google announced it on 11 January 2026, developed with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by more than 20 companies including Stripe, Visa and Mastercard.
It covers the whole shopping trip: searching a catalogue, building a cart, checking out, linking a customer account and order updates. A shop publishes a profile at a fixed address on its website that tells agents what it supports.
Google uses UCP so people can buy from eligible businesses inside AI Mode in Google Search and the Gemini app. Drafts for hotel stays and food orders exist, but they are drafts. UCP can run over MCP and A2A, so these standards work together more than they compete.
For a small business: shops, through their platform. Shopify, for instance, connects eligible stores to AI channels such as ChatGPT, Google's AI Mode and Gemini, Microsoft Copilot and Meta, switched on by default.
A few smaller signposts
Agents also look for some simpler files:
- robots.txt: your site's rules for automated readers. If it shuts out AI search readers, nothing else on this page matters.
- llms.txt: a short summary of your business written for AI tools. See What is llms.txt, and does your website need one?
- WebMCP: a proposal that lets a web page tell a browser agent what a form does. It is a draft from a community group, not a web standard, and Chrome is trialling it.
- API catalog and OAuth discovery: finished internet standards (RFC 9727 and RFC 9728) for businesses with their own software interfaces and customer logins. Most small businesses can skip them.
Which ones matter for your kind of business
- BookingsYour booking platform, and a plain "Book" link with a labelled form.
- ShopsUCP and ACP, through your shop platform.
- Enquiries and quotesA clear, labelled enquiry form, plus your phone number as plain text.
- Information onlyrobots.txt first; llms.txt if it is cheap for you.
Bookings (restaurants, salons, clinics, tours). Agents usually book through your booking platform, not your website. Find out whether your booking system connects to AI assistants. On your own site, keep a plain "Book" link on the homepage and make sure every booking form field has a clear label.
Shops. UCP and ACP matter, through your shop platform. Check its AI channel settings, and keep products, prices, stock and policies complete and correct.
Enquiries and quotes (trades, professional services). No assistant has a general route for "ask this business for a quote" yet. A clear, labelled enquiry form, plus your phone number as plain text, is what helps today. WebMCP is the one to watch.
Information only. robots.txt first; llms.txt if it is cheap for you.
What "agent-ready" means
"Agent-ready" is a loose term for a website that agents can read, understand and act on. It comes down to three questions:
- Can AI read your site? Your rules let AI readers in, and your pages carry real text.
- Can agents find what you offer? Your booking system, shop or services are listed in a form machines read.
- Can agents act? Your forms are labelled, your platform's AI connections are on, and your shop speaks the shopping standards.
Key fact
No file or setting makes an assistant choose your business. Anyone who promises that is guessing.
What you can do now
- Check robots.txt. Make sure it does not block AI search readers such as OAI-SearchBot (ChatGPT search) by accident.
- Ask your platforms. Wix, Shopify and many booking systems already do part of this work for you. Switch on what they offer.
- Make the main action easy to find. A plain "Book", "Order" or "Get a quote" link on the homepage, and a form with labelled fields.
- Keep your facts the same everywhere. Name, phone, hours and services should match on your website and your listings.
- Do not fake it. Never publish a shop profile or connection file for a service you do not have. A signpost that points at nothing is worse than none.
- Look again in a few months. Most of these standards are drafts and change often.
Our agent-ready audit checks these signposts on your website and gives you plain steps for anything missing.
Honest about the timing
Adoption is early. In April 2026 Cloudflare scanned the 200,000 most visited websites: files that announce an MCP server or a software interface to agents appeared on fewer than 15 of them, and only 4% had declared their AI usage preferences in robots.txt. Its own answer to "how agent-ready is the web?" was "not very".
Tip
So treat this as cheap preparation, not a sales channel. Do the inexpensive steps, rely on your platforms for the heavy ones, and expect nothing from agents until customers use them for businesses like yours.
Sources
- Introducing the Model Context Protocol, Anthropic, read 13 November 2026
- MCP joins the Agentic AI Foundation, Model Context Protocol blog, read 13 November 2026
- MCP specification, Model Context Protocol, read 13 November 2026
- About the Wix Site MCP, Wix Developers, read 13 November 2026
- Announcing the Agent2Agent Protocol (A2A), Google for Developers, read 13 November 2026
- Linux Foundation launches the Agent2Agent Protocol project, Linux Foundation, read 13 November 2026
- A2A joins AAIF, Agentic AI Foundation, read 13 November 2026
- A2A specification, A2A Project, read 13 November 2026
- Agentic Commerce Protocol, OpenAI and Stripe, read 13 November 2026
- Agentic Commerce Protocol repository, OpenAI and Stripe on GitHub, read 13 November 2026
- What it means that the leader in agentic commerce just pulled back, Forrester, read 13 November 2026
- Under the hood: Universal Commerce Protocol, Google for Developers, read 13 November 2026
- Universal Commerce Protocol, UCP Authors, read 13 November 2026
- UCP specification overview, UCP Authors, read 13 November 2026
- Agentic storefronts, Shopify Help Center, read 13 November 2026
- WebMCP draft community group report, W3C Web Machine Learning Community Group, read 13 November 2026
- WebMCP, Chrome for Developers, read 13 November 2026
- RFC 9727: api-catalog, IETF, read 13 November 2026
- RFC 9728: OAuth 2.0 Protected Resource Metadata, IETF, read 13 November 2026
- Overview of OpenAI crawlers, OpenAI, read 13 November 2026
- Introducing the Agent Readiness score, Cloudflare, read 13 November 2026


