Most articles about Claude for marketing are about writing. Claude drafts ad copy, rewrites a landing page, summarises a brief. That part works and it has worked for two years. What changed in 2026 is that Claude can be connected to the ad accounts themselves, read yesterday's numbers, and, with the right connector, change them.
This post covers what that looks like for a media buyer running Meta and Google Ads. What Claude does with no connection, what a custom connector adds, which plans allow one, where a review step belongs, and how Claude compares with ChatGPT and Perplexity for the same job. I run Meta and Google accounts every week and I build Adsap, which is one of the connectors described below. Every claim has a source at the end.
What Claude does with no connection at all
Out of the box, Claude has no access to your accounts. It can still do the writing half of the job: primary text and headline variations, a creative brief from a product page, a first pass at audience angles, or an explanation of why an ad was rejected.
The limit is data. Paste a spend report into the chat and Claude reads it well. Ask what happened in your account last week and it cannot know. Anything it says about your numbers without a connection is a guess, and a confident one.
Connectors: the part that changed
A custom connector points Claude at a remote server that exposes tools it can call. The protocol is MCP, the Model Context Protocol. Once connected, Claude calls those tools during a conversation and answers from live data rather than from what you pasted.
Anthropic's plan coverage is wide. Its help pages say that custom connectors using remote MCP "are available on Claude, Cowork, and Claude Desktop for users on Free, Pro, Max, Team, and Enterprise plans. Free users are limited to one custom connector." On Pro and Max you add one under Customize > Connectors, then "Add custom connector" and the server URL. On Team and Enterprise, only Owners can add a connector to the organisation, and each member enables it for themselves.
That Free line matters more than it looks. A media buyer can test an ads connector on a free account before paying for anything, which is not true everywhere else.
The questions media buyers actually ask
Once an ads connector is attached, the useful prompts are boring and specific. These come up most often in my own accounts.
What changed in act_123 in the last seven days, and which ad sets are spending with no conversions?
Show ROAS by campaign for the last 14 days on both my Meta and my Google account, and flag anything under 2.5.
Which active ads have the worst frequency this week, and what are they spending?
None of those is a launch. The first value of an assistant on an ad account is reading: pulling numbers you would have opened four tabs for, and joining two platforms in one answer. Reads are also the safe half, because a read cannot change a client's account.
Reads are safe, writes need a review step
The moment a connector can write, the risk changes. A mis-parsed sentence can pause a campaign that pays the bills, or launch an ad nobody approved. Anthropic's own guidance is to connect only servers you trust, and the same caution belongs on the server side.
The pattern that works is a preview before every write. The assistant proposes the change, the change is rendered in full for you to read, and nothing reaches the platform until you confirm. Anything created is created paused, so a mistaken launch spends nothing while you look at it.
You ask
Launch three ads for the spring set.
Preview
Every field rendered before anything is sent.
You confirm
Nothing reaches the platform until you do.
Created paused
A mistake spends nothing while you look.
Three questions are worth asking of any ads connector before you let it write:
- Does every write show a preview you can read before it happens?
- Is a new campaign, ad set or ad created paused by default?
- Is there a log of what the assistant did, after the fact?
If the answer to any of those is no, keep the connector on reads.
Claude, ChatGPT and Perplexity for the same job
All three assistants hold custom connectors now. What differs is the plan you need, and whether the connector is allowed to write.
Claude
Custom connectors
Free, Pro, Max, Team, Enterprise
Free holds one connector
ChatGPT
Custom connectors
Business, Enterprise, Edu
Pro reads only, web only
Perplexity
Custom connectors
Pro, Max, Enterprise
Enterprise admins can share one
| Custom connectors | Write actions | Notes | |
|---|---|---|---|
| Claude | Free, Pro, Max, Team, Enterprise | Yes | Free is limited to one connector. Owners add them on Team and Enterprise. |
| ChatGPT | Business, Enterprise and Edu for full MCP; Pro for read and fetch in developer mode | Business, Enterprise and Edu only | Web only. An admin enables developer mode and publishes the app. |
| Perplexity | Pro, Max, Enterprise | Yes | Enterprise admins can share a connector across the organisation. |
The practical reading: Claude is the cheapest place to try an ads connector, and the only one where a free account can hold one. ChatGPT is the most locked down, because write actions sit behind a Business plan and an admin who publishes the app. Perplexity sits between the two.
What Claude will not do for you
Three honest limits, because the category oversells itself.
It does not replace judgement on creative. Claude writes serviceable variations quickly. Deciding which angle deserves a real budget is still your call.
It does not see what it has no tool for. An assistant answers from the tools its connector exposes. If the connector has no tool for Instagram media, then Claude cannot see your Instagram posts, however the question is phrased.
It does not make a bad account structure good. An assistant that reads a messy account faster gives you messy answers faster.
Running Meta and Google Ads in Claude with Adsap
Adsap is a connector built for this job. It exposes more than 90 tools across Meta Ads and Google Ads to the assistant you already use, so one conversation covers both platforms. Setup follows the custom connector path above, and the AI copilot setup guide has the click path for Claude, ChatGPT and Perplexity.
The guardrails match the three questions above. Every create is previewed in a card you confirm, everything created is paused, and each action is recorded. The full list is in the safety model. For the reading half only, the copilot answers performance questions without touching anything:
Which campaigns need my attention across both platforms this week, and how much is at risk?
Meta ships its own official connector, which is free and good at reporting. The row-by-row comparison against Adsap is on Adsap vs Meta's official Ads AI connector, and the account setup is in the Meta Ads integration docs.
Frequently asked questions
Not on its own. Claude connects through a custom connector, a remote MCP server that exposes tools for those platforms. Once one is attached, Claude can read account data and, depending on the connector, change it.
No. Anthropic's help pages say custom connectors are available on Free, Pro, Max, Team and Enterprise plans, and that Free users are limited to one custom connector.
Only with a review step. Look for a connector that previews every write before it happens, creates campaigns and ads paused, and keeps a log of what was done. Keep a connector on reads until you have seen those three things work.
For this job it is more accessible. Claude allows a custom connector on every plan, including Free. ChatGPT limits full MCP with write actions to Business, Enterprise and Edu plans, on the web only.
Sources
- Claude Help Center, Get started with custom connectors using remote MCP: "Custom connectors using remote MCP are available on Claude, Cowork, and Claude Desktop for users on Free, Pro, Max, Team, and Enterprise plans. Free users are limited to one custom connector." Also the Customize > Connectors path and the Owners-only rule on Team and Enterprise. https://support.claude.com/en/articles/11175166-getting-started-with-custom-connectors-using-remote-mcp
- OpenAI Help Center, Developer mode and MCP apps in ChatGPT, checked 2026-09-17: "Full MCP (Model Context Protocol) support, including modify/write actions, is rolling out in beta to ChatGPT Business, Enterprise, and Edu plans"; "Apps, full MCP support, and developer mode are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web"; MCP apps on mobile, "No - web only"; Pro users "can connect MCPs with read/fetch permissions in developer mode". https://help.openai.com/en/articles/12584461-connectors-in-chatgpt
- Perplexity Help Center, Adding Custom Remote Connectors: custom remote connectors from an MCP server URL for Pro, Max and Enterprise subscribers, with Enterprise admins able to share a connector across the organisation. https://www.perplexity.ai/help-center/en/articles/13915507-adding-custom-remote-connectors
- Model Context Protocol, the open standard behind custom connectors. https://modelcontextprotocol.io
- Adsap docs, AI copilot setup, the safety model and the Meta Ads integration. The tool count is the Adsap connector's own registry at the time of writing.