MCP servers explained for website and marketing teams

An MCP server is a connector that lets Claude Code read from, and sometimes act in, another system: your CMS, your host, your analytics or your CRM. For a website team, the useful ones today are official servers published by the platforms themselves, including WordPress, Shopify, Webflow, Wix, Vercel, Netlify, GitHub, HubSpot, Notion and Figma. Google publishes an experimental, read-only one for Google Analytics, but not for Search Console.

MCP stands for Model Context Protocol. Anthropic's docs call it "an open standard for connecting AI tools to external data sources and services." Without it, Claude Code works on the files in your project folder and the commands in your terminal. With it, Claude can pull last month's traffic for a page, look up a deploy log or draft a post in your CMS from the same session.

That reach is the point and the risk. Every server you connect gives Claude some of your access to that system, and Anthropic says plainly that it "does not security-audit or manage any MCP server." This guide lists the official servers that exist for common web tools as of October 2026, what each can do, and how to connect them without handing over more than you meant to.

What MCP is, in plain terms

Think of an MCP server as a menu of actions another system offers to an AI tool. A Google Analytics server might offer "run a report". A hosting server might offer "list deployments" and "read build logs". Claude Code sees the menu, decides when an action helps with your task, and asks to use it, subject to the same permission system that governs file edits and terminal commands.

Servers come in two kinds:

• Remote servers run on the vendor's infrastructure. You connect with a URL and usually sign in with OAuth, the "log in and approve access" flow you know from other apps.
• Local servers run on your own machine, started by a command. They often use credentials stored on your computer.

MCP is not specific to Claude. The same servers usually work with other AI clients, which is why vendor pages often list Claude Code next to Cursor, Codex or VS Code. Our Claude Code glossary defines MCP and the related terms, such as connectors and plugins.

How Claude Code connects to a server

Anthropic's MCP documentation describes three places a server's configuration can live, called scopes:

Two details matter for teams. First, Claude Code asks each person to approve project-scoped servers from .mcp.json before using them in an interactive session, so a teammate can't silently add a server to everyone's setup. Second, those approval prompts don't appear in non-interactive runs, such as scripts and cloud sessions, which load project servers without asking. Review changes to .mcp.json in pull requests like any other code.

Anthropic also points to its Anthropic Directory at claude.ai/directory, which lists reviewed connectors that use the same MCP infrastructure.

Official MCP servers for the tools web teams use

The table lists servers published by the platform itself, checked on the vendor's own pages on October 14, 2026. Status labels change, so confirm on the vendor page before you rely on one.

Shopify also runs a Storefront Catalog MCP server, but it is for shopping agents that search a store's products, not for building or editing the store.

Which servers are worth connecting first

Start with the job, not the list. Most website teams get the most from three kinds of connection:

• Platform documentation servers. Shopify Dev MCP and WordPress Studio give Claude current, platform-specific knowledge while it writes theme or plugin code. They read docs rather than your live store or site, so the risk is low. Our guides to Claude Code on Shopify and Claude Code on WordPress show where they fit in a workflow.
• Read-only data servers. Google's Analytics MCP server is "available for read requests only", which makes it a sensible first data connection. Pair it with Search Console data pulled through the API, for example in an audit like those in our Claude Code for technical SEO guide.
• Hosting and repository servers. Vercel's server can inspect deployments and GitHub's can work with pull requests and issues, which helps when something breaks. Be aware that these can also act, not just read.

Leave servers that write to live content, such as WordPress.com, Wix and HubSpot with write access, until your team has a review routine that works.

How to add one, step by step

1. Find the vendor's own instructions. Use the setup page on the platform's developer site, not a blog post or a random repository. Most give a ready-made Claude Code command.
2. Add it at local scope first. Run the command in your project, sign in when asked, and try it yourself before sharing it with the team.
3. Check that it connected. Run /mcp in a session to see each server's status.
4. Use the narrowest credentials. Pick read-only scopes, a staging site or a limited account where the vendor offers one.
5. Try a read-only task. Ask for a report or a lookup before anything that changes data.
6. Share it on purpose. When it's proven, move it to project scope in .mcp.json through a reviewed pull request, with any keys referenced from environment variables rather than written into the file.

The Vercel command is a typical example: claude mcp add --transport http vercel https://mcp.vercel.com, followed by a sign-in. Commands change, so copy the current one from the vendor's page.

Security precautions

An MCP server acts with whatever access you give it. Vercel's MCP documentation spells this out: connecting "grants the AI system you're using the same access as your Vercel user account." The same logic applies to every server on the list. These habits keep that manageable:

• Trust the publisher. Anthropic's docs say to "verify you trust each server before connecting it." Prefer servers from the platform's own organization.
• Watch for prompt injection. Prompt injection means hidden instructions inside content Claude reads, such as a web page, a form submission or a CRM note. Anthropic warns that "servers that fetch external content can expose you to prompt injection risk."
• Least privilege. Read-only scopes where they exist, such as the read-only Analytics scope Google's server requires. Anthropic's own database example uses a read-only database user. For GitHub, a fine-grained token limited to the repositories involved.
• Keep secrets out of shared files. Never paste tokens into prompts, CLAUDE.md or a committed .mcp.json.
• Use permission rules. Rules can allow or deny a whole server (mcp__github) or one tool (mcp__github__<tool>). Deny the write tools you don't need.
• Stage before live. Point write-capable servers at a staging site first, and keep a person approving anything that publishes.
• Control it centrally. Organizations can deploy a fixed server set or set allow and deny lists for MCP servers through managed configuration.

Our guide to using Claude Code safely on a live website covers permissions, secrets and review rules in more depth.

When not to use MCP

MCP is not always the best route. If you need a one-off export, downloading a CSV from GA4 or Search Console and giving Claude the file is simpler and gives it no live access. If a platform has a good command-line tool, such as WP-CLI or Shopify CLI, Claude can often use that in a staging environment instead. And if the only server for a tool is an unofficial one from an unknown developer, wait. Our guide to Claude for small-business data analytics shows the file-based approach.

Conclusion

MCP turns Claude Code from a tool that edits files into one that can read your analytics, check your deploys and work in your CMS. Start with official, read-only or documentation servers, connect them at local scope, use the narrowest credentials, and add write access only once you have a review routine. If you want help choosing servers, writing permission rules and setting up managed configuration for a team, see our Claude Code consulting service. More guides are in our Claude Code Guide.

Frequently asked questions