Resources/Comparisons

The best MCP servers for marketing in 2026: a working PPC specialist's stack

Daniel Popa·Founder, AdPlug·September 30, 2026·16 min read

The best MCP servers for marketers in 2026 are the ones that plug the AI you already use into your own tools: the ad accounts, analytics, the tag manager, the CRM, your call notes. My list is AdPlug for Google Ads and LinkedIn Ads, Google's Analytics server, Stape's server for Google Tag Manager, an open-source Search Console server, DataForSEO, HubSpot, PostHog, Granola, Webflow, Tally, Canva, Replicate, and Higgsfield. I've run Google Ads and LinkedIn Ads for B2B companies for 12 years, and AdPlug is my product, so read this as one PPC specialist's working stack and weigh it accordingly. For each one I'll tell you what it can read, what it can change, and what stands between the AI and that change.

Last checked on 30 September 2026 against each vendor's own documentation. The main page I used is linked in each entry, and I'll go through the lot again every month.

The short version

Fourteen servers, grouped by the job they do. If you only read one column, make it the third.

In one line: connect your ad accounts first, then whatever tells you what happened after the click. For a PPC specialist that's AdPlug and Google Analytics. For a B2B SaaS team it's AdPlug and HubSpot. Add the rest the week you need them.

What an MCP server is, if you run marketing

MCP stands for Model Context Protocol. Think of an MCP server as a plug socket for AI. You connect one of your tools once, signing in through that tool's own login screen, and from then on Claude or ChatGPT can run a report in it or pause a campaign because you asked. No exports, no CSVs, no clicking six levels deep to find one number.

It isn't the same thing as an API. An API is what a developer writes code against. An MCP server wraps that API in actions the AI can pick for itself, so nobody has to write anything. It also isn't one more of the AI tools for marketers with its own chat box and its own login. There's nothing new to log into. It makes the AI you already pay for useful on your own data.

How I chose, and what I haven't tested

I kept a server if it's published by the company whose product it connects to, or by someone with an obvious reason to keep it working, and if its docs say plainly what it reads and what it changes. Search Console is the exception: Google has no documented server for it, so I picked an open-source one. Then I cut anything that doesn't come up in a normal week of running paid media for B2B companies.

AdPlug is mine, and my client accounts run through it. Seven of the others are connected to my own Claude: Stape's Tag Manager server, DataForSEO, PostHog, HubSpot, Granola, Webflow, and Replicate. Where I give a tool count for those, it's what the connector showed on 30 September. I haven't connected Google Analytics, Search Console, Tally, Canva, or Higgsfield, so those entries come from the docs. I didn't query a single client account to write this.

MCP servers for PPC specialists

I love the strategy side of this job and I hate the admin. My working day on an account is a short list of checks. Read the search terms and add the negatives. Find the keyword I've been missing for months. Kill the Demand Gen placements eating budget. Split the LinkedIn leads by job title and seniority. Flag who's underspending by Friday. Check every setting after a launch. That last one sounds dull until search partners is switched on and nobody noticed (it has been me). Every server in this section either takes one of those checks off my hands or tells me what happened after the click.

AdPlug for Google Ads

AdPlug is the connector I built for that list, so read this entry as the maker's description. On Google Ads it gives the AI 16 tools. For reporting, that covers every account under a manager login, account and campaign performance, daily trends, conversions by action, search terms, shared negative lists, keyword ideas, and custom queries for anything else.

It writes as well as it reads. Changes go through a preview tool and an execute tool (negative keywords have a pair of their own), and the main pair covers 149 named operations as of 30 September: creating, pausing, and enabling campaigns, ad groups, and keywords, budgets and bidding strategies, placement exclusions, responsive search ads and assets, Performance Max asset groups, Demand Gen ads, audiences and remarketing lists, conversion actions, offline conversion uploads, labels, and experiments.

Nothing changes on a request alone. The AI has to call the preview tool first, which returns the exact operation and the values it's about to set. The execute tool refuses to run unless a preview is waiting, and on the main pair a preview left for fifteen minutes is thrown away. One batch can carry up to 100 operations, previewed as a single list. It shows me what it's about to change before it changes it, which is the part that lets me sleep.

Pull the search terms for the last 30 days in [account], sorted by cost. List the ones with spend and no conversions, group them by theme, and preview the negatives you would add at campaign level. Do not execute until I say so.

Search terms are step two of the 30-minute audit. The other five steps each come with a prompt.

AdPlug for LinkedIn Ads

LinkedIn doesn't have an MCP server of its own (with most AI ads tools it's Google first, Meta second, LinkedIn never). The nearest thing is LinkedIn's plugin for ChatGPT, in beta, which reads ad set performance and changes nothing. AdPlug's LinkedIn Ads server has 10 tools. Reports come in seven types, and the one I'd start with is demographics: spend and conversions by job title, seniority, company size, and industry. In Campaign Manager that's the demographics tab, one breakdown at a time, then an export and a pivot table to make it readable. There's also a targeting lookup with audience counts, so you can see how big an audience is before you build it, and a search of LinkedIn's public Ad Library that needs no ad account, only a connected LinkedIn login.

Writes are 25 operations: campaign groups, campaigns, ads in nine formats including Thought Leader Ads, audiences and list uploads, lead gen forms, and conversion rules. A separate tool sends conversion events in bulk. There's a preview tool here too, but on LinkedIn the server doesn't force it, so ask for the preview in your prompt every time.

Every Monday I ask it for cost by job seniority over the last 90 days. The fuller version:

Cost per conversion by job seniority for the last 90 days in [account], then the same by job function. Tell me which segments are spending without converting.

There are more like it, for both platforms, in the prompt playbooks.

Both platforms run on one AdPlug account. The free plan is view only on one ad platform, with as many ad accounts as you like, 500 requests in the first month and 50 a month after that. Pro is $29 a month for both platforms, with 5,000 requests and 200 edits. Max is $89 with no cap on either, plus three team members. To connect, you create an account, link the ad platform through Google's or LinkedIn's own sign-in screen, and paste one address into your AI: https://api.adplug.app/mcp/google-ads or https://api.adplug.app/mcp/linkedin-ads. The Google Ads setup guide and the LinkedIn Ads one take a few minutes each, and pricing has the detail.

How that compares with the platforms' own connectors

Google publishes an official Google Ads MCP server. Its guide lists the mode as read-only in the current release, and you run it yourself, on your own machine or in your own Google Cloud project. LinkedIn's plugin lives inside ChatGPT. Here's what each one does, taken from the vendor's own page.

Meta is the odd one out. It hosts its own ads MCP server, in open beta since April, which reads reports and also creates and edits campaigns, with new ones starting paused. For Meta, I'd start there.

Google Analytics MCP

Analytics shows what people did after the click, which the ad platform only half sees. Google's Analytics MCP server is official, open source, and labelled experimental in its GitHub repository. It lists your accounts and properties, shows which Google Ads accounts are linked to a property, runs standard, funnel, and realtime reports, and returns your custom dimensions and metrics. Google's guide says it's available for read requests only, so it can't edit a property.

The catch is the setup. It runs on your own machine through a command-line tool called pipx, and it needs a Google Cloud project and credentials with the Analytics read-only scope. If you're comfortable in a terminal, that's an afternoon. If words like "pipx" mean nothing to you, hand it to whoever does technical things at your company. It's worth it for questions like this one:

For the last 28 days, show sessions, engaged sessions, and key events for traffic where the session source / medium is google / cpc, by landing page. Flag landing pages that get paid traffic and no key events, and say which have too little traffic to judge.

Google Tag Manager MCP, by Stape

Most tracking fixes mean opening Tag Manager, and that's slow work by hand. Stape's Tag Manager server is open source and comes in two flavours: a hosted one at https://gtm-mcp.stape.ai/mcp with Google sign-in, and a local one that runs on your own credentials. Stape counts the same 18 tools in both (the hosted connector shows 19 in my Claude, and the extra one clears your sign-in). They cover tags, triggers, variables, workspaces, versions, and permissions. It can create, update, and remove tags, triggers, and variables, and it can publish a version.

I used the hosted one to add consent settings to AdPlug's own container, and three things from that are worth passing on. It reaches every container your Google login can open, client ones included, so I make the AI read the container ID back to me before it writes anything. An update replaces the whole tag or trigger, so the AI has to fetch the current one first and send all of it back (Stape's changelog warns that any field left out is deleted). And Google's Tag Manager API quota is tiny, a quarter of a request per second per project. I hit a quota error after two or three calls and had to wait a minute each time.

Stape's own advice, in a post on controlling AI in Tag Manager, is not to sign in as yourself, because on most client containers that hands the AI your publishing rights. Use a separate Google account with Edit rights and nothing higher, added only to the container it needs.

Search Console MCP

Google hasn't documented a Search Console server the way it has for Analytics. What exists is community work. The one I'd look at is AminForou/mcp-gsc: open source under an MIT licence, about 1,800 stars on GitHub, 20 tools, installed on your own machine with Google credentials you create yourself. It reads search analytics and URL inspection data, and it can submit a sitemap. The tools that add or delete a site, or delete a sitemap, are switched off until you turn them on. For a PPC specialist the use is narrow but good: finding the queries you already rank first for while you're paying for the click.

DataForSEO MCP

The keyword research behind this blog runs through DataForSEO's MCP server. It's official, hosted at https://mcp.dataforseo.com/v3/mcp or run on your own machine, and small: four tools, three that look things up in DataForSEO's documentation and one that makes the request. What comes back is data: search volumes, keyword difficulty, live search results for a query in a given country, backlinks, and what AI assistants answer for a prompt. There's no marketing account behind it to change. Every request comes out of your DataForSEO balance, pay as you go, with a $50 minimum payment and no extra charge for the MCP connection. There's no subscription, so it's a cheap way to check whether a term is worth a campaign before you build one.

MCP servers for B2B SaaS marketing teams

In B2B SaaS the ad platform counts a form fill as the conversion, and the team is judged on pipeline. So this stack joins the ad account to what happened weeks later. HubSpot holds the deal, PostHog shows what the signup did in the product, and Granola has what the buyer said on the call.

HubSpot

If you're a marketer, the way in is HubSpot's connector for Claude, which HubSpot lists as available on all its products and plans. You'll need a paid Claude plan. Super admins and anyone with App Marketplace access can connect it straight away, and everyone else needs a super admin to approve it first. Behind it sits HubSpot's remote MCP server, which also powers its connectors for ChatGPT, Copilot, and Gemini. The connector in my Claude shows 30 tools.

It reads contacts, companies, deals, tickets, campaigns, and marketing email performance. It creates and updates CRM records, landing pages, blog posts, and marketing emails, and it can't delete anything. For record changes, HubSpot's documentation says the tool shows a summary of the proposed changes and needs explicit confirmation before it runs. Bulk writes are capped at ten records at a time (I'd call that a feature), and every change lands in HubSpot's audit log against both you and the connector.

One thing to know before you trust a number from it: the campaign attribution report only covers closed-won deals. Ask for that when you mean revenue, and ask for campaign data when you mean everything still in the pipeline.

List deals created in the last 90 days where the original source is paid search or paid social. Group them by stage and show amount and days open. Read only, do not change any record.

PostHog MCP

AdPlug's own product analytics live in PostHog, and I query them through its MCP server. The three articles linked in this site's footer got there that way: I asked which ones bring in the most visits from organic search. It's PostHog's own server, hosted at https://mcp.posthog.com/mcp, and free to connect, with the usual PostHog billing for the data your queries touch. It can write as well as read, feature flags and experiments included. If you only want answers, PostHog documents a read-only mode: add ?readonly=true to the address and every tool that creates, updates, or deletes disappears.

Granola MCP

Call notes are a good source of ad copy, because they hold the words buyers use. Granola's MCP server puts your meeting notes within reach of the AI. It's Granola's own, hosted at https://mcp.granola.ai/mcp, and you sign in through the browser. There are six tools and every one of them reads: list meetings, fetch the notes, fetch a transcript, ask a question across the lot. The free plan covers your personal notes from the last 30 days, and Granola says some folder, search, and transcript tools are paid only. The transcript is where the exact quotes are, so that one matters. It follows whichever workspace is active in the Granola app and won't combine notes across workspaces.

Go through my customer calls from the last 30 days. Pull the exact phrases people used to describe the problem they had before buying, with the meeting each came from. Then draft five LinkedIn ad headlines built on those phrases.

Webflow MCP

The landing page is the next place a campaign can fall over. Webflow's MCP server is available on every site plan, including the free Starter one. Webflow's help article says anyone with a site role other than Reviewer can connect it and that it respects that person's permissions, and its developer docs say each sign-in covers one workspace. The connector in my Claude lists 34 tools. They create and edit pages, components, and CMS items, and they can publish a site or a single page, so treat it like a teammate with publishing rights. The analysis side is thin for now: it needs the Analyze add-on, and Webflow says it reads site-level performance only.

Tally MCP

For the form itself, Tally's MCP server is free on all plans, including the free one, at https://api.tally.so/mcp. There's also a ready-made connector for Claude and a plugin for ChatGPT. Tally's developer page still labels the server as beta. It builds forms from a description, edits existing ones, and fetches submissions. It can't delete a form or a submission, and a new form stays a private draft until you ask for it to be published. Fetching submissions is the useful bit for a marketer. Point it at the open text field on your demo request form and ask what people say they want.

MCP servers for ad creative: Canva, Replicate, and Higgsfield

These three make the images and video that go into the ads. None of them can touch an ad account.

Canva MCP

If your team already keeps its templates in Canva, start here. Canva's MCP server is official, at https://mcp.canva.com/mcp, and open to every Canva account, free ones included. A connected AI can search your designs, generate new ones, edit them, and export files. Brand templates need a paid Canva plan, and Canva says generating designs may use your monthly AI credits. On a Canva team, an admin switches the connector on or off for everyone at once. I haven't connected this one myself.

Replicate MCP

Replicate's MCP server gives the AI Replicate's catalogue of image and video models. It's Replicate's own, and comes as a hosted server (the one Replicate recommends) or a local one you run yourself with an API token. The connector in my Claude lists 29 tools, for searching models, reading what inputs one takes, starting a run, and fetching the result. There's nothing in a marketing account for it to change, but every run costs money. Replicate's pricing page says most models are billed by the time they take to run, and some by input and output, per image for example.

Higgsfield MCP

Higgsfield's MCP server does the same job for its own image and video models, at https://mcp.higgsfield.ai/mcp. In Claude it's a custom connector you add by address, and in ChatGPT it's a plugin from the directory. Its help centre says you need an active paid subscription and that every generation through a connected AI uses credits. There's no built-in cap on them, so Higgsfield's own advice is to tell the AI how many credits it may spend and to ask before it generates anything. I haven't connected this one either.

Whichever you use, look at what comes out before it goes anywhere near an ad account, above all anything that shows a real product or a real person.

Before you let one of them change anything

If you manage other people's money, read this bit.

• Start every server on read. AdPlug's free plan is view only. PostHog has a read-only switch. Google's Analytics server can't write at all. Stay there until you've had a week of answers you've checked.
• Know whose permissions it's borrowing. A connector acts as the login you gave it. Mine can see every client ad account and every client Tag Manager container I can, which is why the account or container name goes into the prompt and gets read back to me before anything changes.
• Find out what stands between the request and the change. AdPlug's Google Ads server won't execute without a preview. HubSpot's tool is documented to show the proposed changes and ask for confirmation. Tag Manager and Webflow can publish, so the real control there is the permission of the user you connect. An instruction in a prompt is only a request.
• Keep write tools on ask-every-time. Claude lets you set a connector's write tools to need approval, and ChatGPT has an Always ask setting for each app. Leave both on for anything that spends money or publishes.
• Put a number on anything that charges per run. DataForSEO, Replicate, and Higgsfield bill by the request or the generation. Tell the AI how many it's allowed.
• Test on something you own. When I try a new write operation, it's on my own account, on a draft campaign named DELETE-ME. Never on a client's.

Which to connect first, and why not all of them

Two or three at a time. Every server you connect adds its tools to the pile the AI has to choose from, and Anthropic's own developer documentation says Claude's ability to pick the right tool degrades once you go past 30 to 50 available tools. HubSpot's connector is 30 on its own. Webflow's is 34 and Replicate's is 29. Add AdPlug's 16 for Google Ads and that's 109 tools in front of the AI before you've asked it anything. The assistants are getting better at searching a long list, but a short list is still the easier fix. Switch on what the task needs.

• PPC specialist, client accounts: AdPlug for the platforms you run, then Google Analytics if someone can install it for you. Add the Tag Manager server on the day you have a tracking job.
• B2B SaaS team: AdPlug for LinkedIn Ads, then HubSpot. Add Granola when you're writing new ads and PostHog when the question is what signups did next.
• Creative week: switch the reporting servers off and one design or image server on.

Which AI assistants these work in

Any assistant that accepts a remote MCP address. In AdPlug's own logs for the last 90 days, most people connect from Claude, then ChatGPT, Claude Code, Codex, and Cursor. I've also connected it to Grok, Grok Bot, Hermes Agent, and OpenAI's Dots, and each has a setup guide: Grok Bot, Hermes Agent, and Dots. Perplexity documents custom remote connectors under its Enterprise features. I haven't tested that one, so I'm going by its docs.

ChatGPT is the one with a catch. Custom servers go through Developer mode, and OpenAI's own pages disagree on whether individual plans can run write actions. I've set out both versions in the ChatGPT setup guide. The two servers you run on your own machine, Google Analytics and the open-source Search Console one, work in desktop and command-line clients such as Claude Desktop, Claude Code, and Cursor. A browser tab can't reach them.

What I left out

Other ads connectors are in the two comparisons linked above, where you can look at them side by side. Notion and Slack both publish official servers. They're where a report ends up, so I'd add them once the reporting itself works. Email platforms and SEO suites have their own too, Klaviyo and Ahrefs among them, and both sit outside the jobs this page is built around. I also left out the hubs that put hundreds of apps behind one connector. This list depends on knowing what each action does to an account, and I can't vouch for a catalogue that size action by action.

If you try one thing from this page, make it a read-only question about your own spend. Connect Google Ads or LinkedIn Ads on the free plan, ask for the search terms or the seniority breakdown, and judge the whole idea on your own account.

P.S. The hours this gives me back mostly go on tennis. My backhand is still a work in progress.