What is the Model Context Protocol (MCP), and why did WordPress add an adapter?

The Model Context Protocol (MCP) is an open standard — released by Anthropic in November 2024 and since adopted by OpenAI and Google — that lets any AI assistant plug into any tool or data source through one shared interface, the way a single USB-C port fits every device. WordPress shipped an official MCP adapter in early 2026 for one reason: so that plug reaches your site, letting an AI agent create posts, update products and read orders by speaking the protocol instead of clicking through wp-admin.

There is a catch we keep having to untangle for store owners, and it is exactly why we watch this closely from the other side of the fence while building Contexta for WooCommerce. The MCP adapter is a door for your agent to run the store — it is not a door for a shopper's agent to find and buy from it. Those are two different protocols solving two different problems, and treating them as one is the most common mistake in the "MCP for WooCommerce" coverage we have read this year.

What is the Model Context Protocol, in plain terms?

MCP is a single, open way for AI clients to reach outside tools and data, so an assistant needs one integration instead of a custom connector per app. Anthropic's own description of it is a "USB-C port for AI applications": before the standard, every assistant-to-tool link was bespoke, so ten clients wiring into ten tools meant a hundred one-off integrations; MCP turns that N-times-M problem into N-plus-M.

It works client-and-server. An MCP host — Claude Desktop, Cursor, an IDE, or an agent you build yourself — connects to MCP servers that expose three kinds of thing:

• Tools: executable functions the agent can call to do something (create a post, update a price).
• Resources: passive, read-only data the agent can pull for context (a settings value, an order record).
• Prompts: reusable templates that guide a specific workflow.

Anthropic open-sourced MCP in November 2024, OpenAI added support across its products in March 2025, and Google followed the same year. That cross-vendor buy-in is the whole point: a tool that speaks MCP is reachable by every major assistant at once, not just the one it was built for.

Why did WordPress add an MCP adapter?

WordPress added the adapter to expose its own functionality as MCP tools and resources, so any AI client can discover and run site tasks without a custom bridge per assistant. The groundwork landed first: the Abilities API arrived in WordPress core 6.9 in November 2025, giving plugins and themes a standard way to register a discrete capability with a permission check and typed input and output. The MCP adapter — part of the official "AI Building Blocks for WordPress" initiative, introduced on the WordPress Developer Blog in February 2026 — wraps those registered abilities as MCP primitives.

Note what this is not. It runs the store, it does not sell through one; the selling path, where a shopper's assistant completes a purchase, is a separate build covered in selling WooCommerce products through ChatGPT.

The connection runs two ways: STDIO for local development through WP-CLI, or HTTP for a remote site through the @automattic/mcp-wordpress-remote proxy, which translates an assistant's MCP calls into the site's REST endpoints. The adapter code moved from Automattic's original wordpress-mcp repository to the shared WordPress/mcp-adapter project, and as of mid-2026 it still sits at a 0.x version — a real, working bridge, but an early one.

Will the WordPress MCP adapter make shopping agents find and buy my products?

No — and clearing up this confusion is the single most useful thing to do before you install anything. MCP is an inbound control plane for an agent you authenticate and trust: it lets your assistant operate your admin. It does nothing to change whether a shopper's ChatGPT or Gemini can discover your product, trust its price, or complete a checkout. That outward-facing job belongs to a completely separate layer — the agentic-commerce protocols mapped out in ACP, AP2, UCP and x402, which is where product discovery and payment actually happen.

Point the MCP adapter at your store and a shopper's assistant is no closer to buying from you than before, because it never touches that endpoint. Read the adapter as a back-office tool, not a storefront, and every decision about it gets simpler.

What can an AI agent actually do to a WooCommerce store through MCP?

Through the WooCommerce MCP integration — a developer preview as of mid-2026 — an authenticated agent can query, create, update and delete products, and query orders, change their status and add order notes. Everything runs as an admin operation behind WooCommerce REST API keys, so an agent can only do what the credentials you hand it are allowed to do.

The practical shape of it:

• Server endpoint: https://yourstore.com/wp-json/woocommerce/mcp, enabled behind an mcp_integration feature flag, HTTPS required.
• Auth: a WooCommerce REST key passed as consumer_key:consumer_secret in an X-MCP-API-Key header.
• Good fits today: bulk product edits, catalog cleanup, migration, and reporting driven by conversation — "set every out-of-stock product in this category to draft" without touching the dashboard.

This is strictly the operator's side. The buyer's side — a customer's agent completing a purchase — never reaches this endpoint. Because it is a preview, expect the tool list and payloads to shift, and pin the version you test against. The exact seven abilities, which of them read versus write, and the ones worth keeping a human in the loop for are broken down in what an AI agent can and shouldn't do with your WooCommerce store.

Is it safe to let an AI agent run your store over MCP?

Only with the same discipline you would give any admin account, because an MCP tool that can delete a product is exactly as dangerous as a human who can. The adapter's power comes entirely from the abilities you expose and the user you connect as, which is where the real failure modes live — an over-broad permission callback or an admin-level key handed to an agent that then acts on a bad instruction.

The honest summary: the security model is sound, but it is your job to configure, not a default. The adapter ships the plumbing; the blast radius is set by the capabilities and credentials you choose to attach. That configuration cost is also why MCP isn't automatically the interface to reach for first — weighed against product feeds and llms.txt by cost and risk, it earns its place only when running the store, not selling through it, is the real bottleneck.

Should a store owner install the MCP adapter today?

It depends entirely on what you are trying to do, and for most stores the honest answer is "not for the reason you first assumed." If your goal is to automate your own back office — bulk edits, content operations, reporting through an assistant you already use — the adapter is genuinely useful and worth trying on a staging site. If your goal is to be discovered and bought by AI shopping agents, it does nothing, and the hours are better spent elsewhere. That second goal is a discovery-and-data question, closer to whether your WooCommerce store needs a UCP-style path at all than to anything in the admin.

Weighed against effort, the adapter earns its place for operators who live in an AI client all day and want to skip the dashboard. For everyone else it is a promising 0.x tool to watch, not a launch-week install.

If MCP won't get you found, what actually will?

The thing that decides whether a shopping agent picks your product is the same unglamorous layer every commerce protocol assumes is already clean: complete, machine-readable product data. An agent has to identify your item, trust its price and confirm it is in stock before any checkout path has anything to execute, which is precisely the record an AI shopping agent needs from each product — a resolvable identifier, an honest price, real availability, content it can read without running JavaScript.

Checking that by hand across a few hundred SKUs is the task nobody finishes, which is why Contexta's Commerce Readiness audit scans every WooCommerce product for the fields agents actually use — image, price, GTIN, brand, weight — and returns a per-product fix link. The MCP adapter is a fine way to let your assistant run the store; getting bought by someone else's assistant is a different door, and it opens on your data, not your admin.