MCP Servers for Marketing: 11 Compared (2026)

Most marketing teams now work with an AI assistant, and most of them use it the same way: copy numbers out of Google Ads, attach a shop export, paste last week's email report, then ask. The answer is only as good as whatever someone gathered beforehand, and it is out of date by the next morning.

MCP servers remove that step. Instead of pasting data into a chat, you connect Claude, ChatGPT or another client straight to the tool, and the assistant fetches what it needs for the question. Since spring 2026 most major marketing tools have shipped a server of their own. They differ a lot, though. Some only read, others create campaigns or send emails. Some connect in three minutes, others need a Google Cloud project.

This guide covers eleven servers that matter for e-commerce and retail teams, each with origin, status, permissions and requirements. We checked every detail on the vendors' own pages on 30 October 2026.

What an MCP server does for marketing

MCP stands for Model Context Protocol, an open standard Anthropic introduced in late 2024 that ChatGPT, Gemini and Microsoft Copilot now support as well. An MCP server is the tool's side of the connection: it declares which functions an AI client may call, such as "run report" or "create campaign", and executes them with your permissions. The client picks the function it needs per question and shows you each call.

In practice, a question like "Which campaigns underperformed last week compared with last month, and what did competitors send in that window?" gets answered from two sources by the assistant itself, instead of someone building two exports.

Three things decide whether a server is useful day to day:

• Origin: an official server comes from the vendor. Community servers are built by someone else on top of the vendor's API. They can be good, but nobody promises they survive the next API change.
• Permissions: does it only read, or can it create, change and send? And can you switch writing off?
• Setup: a hosted server with sign-in (OAuth) connects in minutes. A local server needs Python or Node, API credentials and usually someone technical.

The eleven servers at a glance

Reyo MCP: your competitors' campaigns

Our own server first, so you can weigh it accordingly. Reyo MCP gives your AI client your competitors' campaigns and newsletters, each with discount, period, code and creative, plus upcoming occasions and your own Campaign Planner. It can write two things: planner entries and comments. Both only if you tick "may create and comment" when you allow access, and always only inside your own account.

Setup is one address and a sign-in, no key; the guide walks through Claude, ChatGPT and Claude Code. The server is included from the Professional plan. What it doesn't know: your store data and your ad accounts. That is what the servers below are for.

Best for: knowing what the market is doing before you read your own numbers.

Google Ads MCP: official, but read only

Google has published its own server as open source since October 2025. It has three tools: list accessible accounts, run Google Ads Query Language queries, and look up field metadata. That covers questions about cost, clicks, conversions and search terms across every campaign. It cannot change bids, pause campaigns or create assets, as Google's documentation states plainly.

You run it locally with Python or in your own Google Cloud project; a Google-hosted version is so far a pilot for allowlisted accounts. One hurdle is gone: since 9 September 2026 the Google Ads API no longer needs a developer token, according to Google, and access is tied to the Cloud project instead. Expect an afternoon with someone technical all the same.

Best for: weekly reports and digging into search terms without an export.

Meta Ads MCP: it writes, but launches nothing on its own

Meta opened its Ads MCP Server as an open beta on 29 April 2026. It is hosted, and you sign in with Facebook Login for Business. Beyond reporting, it can create and edit campaigns, ad sets and ads, manage audiences and read test results.

Two safeguards are built in. Anything new is created paused, and only a separate call activates it and starts spend. Since July 2026, business portfolio admins can also set rules on what agents may change, budgets for example. Meta notes that not every account sees every tool yet.

Best for: analysis across many campaigns, and drafts you review in Ads Manager and launch yourself.

Google Analytics MCP: reports inside the chat

GA4 has an official Google server too, which Google labels experimental on GitHub. It pulls account summaries and property details and runs standard, realtime, funnel and conversion reports, custom dimensions included. It only reads; the single exception is property annotations, and only if you explicitly switch that on.

Setup is similar to Google Ads: local Python, a Google Cloud project and viewer access to the property.

Best for: questions like "Why has conversion rate dropped since Tuesday?", split by channel and device, without anyone building an exploration.

Shopify MCP: talk to your store

Search for "Shopify MCP" and you mostly find developer servers. The one meant for marketing teams is the Shopify connector for AI tools, which you approve in Shopify admin. It reads orders, products and analytics. It can only change things once you approve write access, and then only within the permissions of the staff account it acts through.

With write access it manages products, collections, inventory and percentage-off discount codes. Refunds, cancellations, gift cards, the live theme and fixed-amount discounts stay blocked, and it works with one store at a time.

Best for: assortment and discount questions grounded in real sales, such as which category was restocked after the last promotion.

Shopware MCP: in the core, still experimental

Since version 6.7.11 (June 2026) Shopware ships an MCP server in its core, without a feature flag since 6.7.14. It runs on your own store and authenticates with integration keys or OAuth. It searches, reads, creates and updates records, uploads media and sets order states.

One safeguard is well chosen: every write tool runs as a dry run by default, and nothing is saved until that is explicitly turned off. Shopware calls the server experimental until version 6.8.

Best for: teams with their own Shopware instance and a developer in-house. For a pure marketing team it is early days.

HubSpot MCP: CRM and marketing email, no deleting

HubSpot's hosted server has been generally available since 13 April 2026, for every account including Free. It reads CRM data and can create and update contacts, deals, lists, landing pages, blog posts and marketing emails. Campaigns can be created but not edited.

It cannot delete anything, and bulk changes are capped at ten records. Since the fall update in September it also schedules and publishes marketing emails; you have to reconnect once to get the new permissions.

Best for: building segments from a question and dropping email drafts right where they get sent.

Klaviyo MCP: powerful, with writing switched on

Klaviyo's server has been generally available since August 2025, hosted with sign-in or local with a private API key. Its scope is wide: profiles, segments, templates, campaigns, flows and reporting. On the write side it goes as far as creating and sending a campaign. Owner, Admin and Manager roles can use it.

Worth knowing: writing is on by default. The listed Klaviyo connector in Claude and the ChatGPT app cannot be set to read only. If you want that, Klaviyo's docs tell you to add a custom connector with ?read-only=true at the end of the address.

Best for: analysing flows and campaigns. Sending only with a clear approval rule in the team.

Brevo MCP: two routes, two permission levels

Brevo offers two official routes. The Claude connector is free on every Brevo plan. It reads contacts, lists and segments and creates email campaigns and templates as drafts. By Brevo's own account it can never send, schedule or delete.

The developer MCP server, by contrast, uses an API key and has full access, sending included.

Best for: teams that want drafts from the assistant but keep the approval inside Brevo.

Canva MCP: designs from the chat

Canva's hosted server is open to every plan, Free included, and each person signs in with their own account. It generates and edits designs, comments, uploads assets and exports PDF, PNG, JPG, PPTX or MP4. Resizing, brand kits and brand templates need Canva Pro or above.

Best for: turning a campaign brief into first banners and formats that the design team finishes in Canva.

Semrush MCP: SEO data without the export

The Semrush server is hosted with sign-in, with an API key as a fallback. It covers keyword research, backlinks, position data and site audits, plus traffic data with a Trends plan. Projects are read only. The server has no price of its own; queries use API units. Semrush One Starter and Pro+ as well as SEO Classic Pro and Guru include 50,000 units.

Best for: content planning and keyword gaps against competitors.

More official servers

• Amazon Ads: open beta since February 2026, with a separate endpoint for Europe. Updates budgets and creates Sponsored Products campaigns. Open to accounts with active Amazon Ads API access.
• Microsoft Advertising: open pilot since June 2026, read only, free according to Microsoft.
• Ahrefs: hosted, on every paid plan from Lite; queries use API units.
• Similarweb: hosted, requires an API plan.
• Supermetrics and Windsor.ai: one server for many data sources. Both now offer write tools for ad accounts as well, Supermetrics as a beta.
• Salesforce Marketing Cloud Engagement: generally available since June 2026. Creates, publishes and deletes journeys; documented for Claude Code and Gemini CLI.

Where there is no official server yet

• Google Search Console: Google has no server of its own. The most used community server reads performance data and inspects URLs; destructive actions need a separate switch.
• LinkedIn Ads: no official server, and the community projects are small.
• Mailchimp: the announcement of the Mailchimp app for Claude and ChatGPT lists the US, Canada, the UK and Australia.

Read, write, send: the permission question

The list shows a pattern: the closer a server sits to money or to your customers' inboxes, the more it matters what it may do without you. Five rules that have held up for us:

1. Start with reading. Reports and questions are enough for the first weeks. Write access comes once it is clear what it is for.
2. Write only with a stop in between. Good targets have a draft or paused state: paused Meta ads, Brevo drafts, Shopware's dry run, a planner entry. Sending and budget stay with people.
3. Keep outside content away from write access. When an assistant reads competitor newsletters, web pages or reviews, it reads text somebody else wrote. That text can contain instructions aimed at AI assistants (prompt injection). A client that reads outside emails and can send through Klaviyo in the same conversation is a risk. Keep confirmation before every write action switched on.
4. Sort out personal data. CRM and email servers pass customer data to the AI provider. Check with whoever owns data protection which client is approved for that, under which agreement. Aggregated reports answer most analysis questions anyway.
5. Keep a register. Write down which server is connected with which permissions, and disconnect what nobody uses.

Three workflows that work today

The Monday check-in. Reyo answers what competitors sent last week; Google Analytics shows how your own traffic moved by channel. The assistant writes one paragraph for the team meeting. Both servers only read, so the risk is low. How to set up the competitor side without AI is covered in our comparison of competitor tracking tools.

A counter-campaign as a draft. Reyo surfaces a promotion from your main competitor. The assistant proposes a campaign of your own, adds it to the Reyo Campaign Planner after your yes, and creates an email draft through HubSpot or Brevo. Sending happens in the email tool, after your team approves. How such flows are built in general is in the guide to marketing automation.

From brief to format. The planner entry becomes a short brief, Canva turns it into three formats, the design team reviews and adjusts. The assistant saves the blank page here, not the design work.

What the three have in common: data from several sources comes together without anyone exporting. That is the real gain, and more on it is in the guide to marketing data analysis.

Conclusion

Most marketing tools now have an official MCP server, and the big gaps (Search Console, LinkedIn) are shrinking. Choosing is therefore less about what exists than about permissions: which sources should your assistant read, and where may it create something?

A good start is two read-only servers that together answer a question you currently solve with two exports. For your own performance that is usually Google Analytics or Google Ads; for the market, competitor campaigns in Reyo.

All details as of 30 October 2026, checked on the vendors' own pages. MCP servers change month to month, so check scope, status and pricing yourself before you connect one.