Claude connectors for marketers

What a Claude connector actually is

Out of the box, Claude only knows what you paste into the chat and what it learned in training. A connector changes that: it plugs Claude into a live system, such as your Google Drive, your CRM or an ad account, and gives it a set of named actions it can call. Claude decides when to call them, you see each call, and the results land back in the conversation.

Under the hood, a connector is a remote MCP server. MCP, the Model Context Protocol, is the open standard Anthropic published in late 2024 for connecting AI apps to outside tools. The server describes its tools (for example google_ads_search_terms or ga4_traffic_summary), Claude reads those descriptions, and when a question needs live data it makes a tool call.

That matters for marketers because the expensive part of most analysis is not the thinking. It is the exporting, filtering and pasting. A connector removes that step: “which search terms wasted money last month?” becomes a single question instead of a CSV export, a pivot table and a prompt.

Directory connectors vs custom connectors

Claude offers two kinds of connectors. They use the same protocol; the difference is who vetted them and how you add them.

There is also a third, older pattern: local MCP servers that run on your own computer and talk to Claude Desktop. They are powerful for developers but awkward for a marketing team, because every person has to install and update them, and API keys tend to end up in config files. For shared ad accounts, a hosted (remote) connector with per-person sign-in is the safer default.

How to add a marketing connector to Claude

The steps below use MegaMCP as the example because it is what we build, but the flow is the same for any remote MCP server: get the URL, add it, sign in, test it on something harmless.

1. Create an account with the connector's provider.

   For MegaMCP that means starting a free trial: 5 days, no credit card, and a sandbox ad account with fake money.

2. Copy the MCP server URL.

   MegaMCP's is https://megamcp.com/api/mcp. Every AI app uses the same URL.

3. Open Claude's connector settings.

   In Claude on the web or desktop, open the connectors area of your settings and choose to add a custom connector.

4. Paste the URL and name it.

   Call it something short, like MegaMCP. Short names avoid display problems in tool menus.

5. Connect and sign in.

   Claude opens the provider's sign-in page. You pick a workspace and approve access. Claude never sees a password or an API key.

6. Switch it on in a chat and test with a read.

   Ask a question that only reads, such as “Summarize the demo account's last 30 days.” Then try a change on the sandbox and approve it, so you see the whole loop before you touch a real account.

Exact menu names change, so we keep a screenshot-free, current walkthrough at Connect Claude. If you use Claude Code in a terminal, it is one command instead: claude mcp add --transport http megamcp https://megamcp.com/api/mcp. The MCP config generator writes the right snippet for Claude Code, Cursor and VS Code.

What to ask once it's connected

Start with questions you already answer every week. They have a known right answer, so you can check Claude's work.

• Reads data

  Compare last week with the week before for every ad account. Spend, conversions, CPA and ROAS, one row per account.

  Tools it calls: megamcp_cross_channel_summary

• Reads data

  Which Google Ads search terms spent over $50 in the last 30 days without a conversion?

  Tools it calls: google_ads_search_termsads_waste_sweep

• Reads data

  Why did GA4 key events drop last week? Break it down by channel and landing page.

  Tools it calls: ga4_compare_periodsga4_landing_pages

• Reads data

  Audit my 10 latest WordPress posts for missing meta descriptions and long titles.

  Tools it calls: wordpress_seo_audit

For more, the prompt library has copy-ready prompts for each platform, grouped by job, with the exact tools each one calls.

Keeping a marketing connector safe

A connector that can only read is low-risk. One that can change budgets is not, and the risk is not that Claude is malicious. It is that language is ambiguous (“pause the weak ones”), data can contain instructions (a search term or a page title that says “ignore your instructions”), and a confident model can be confidently wrong.

Look for these properties before you give any connector write access to an account that spends money:

• Writes become proposals, not actions. The tool should return a plan with before and after values that a person approves. In MegaMCP every write tool returns a plan; nothing reaches the platform until someone with the right role approves it. See change approval.
• Spend caps the AI can't edit. A per-account maximum daily budget and a maximum percentage increase, enforced by the server, not by a prompt. MegaMCP calls these budget bumpers.
• Undo. Supported changes should record what they need to reverse themselves, and irreversible ones should be labeled before you approve.
• Per-person sign-in and an audit trail. You should be able to see who asked for what, from which AI app.
• Read-only hints. Tools that only read should say so, so Claude can tell a report from an action.

Claude also asks for your permission before it uses a connector's tool, and you can allow a tool once or always. Leave write tools on “ask every time” until you trust the setup.

When a connector doesn't work

• No tools appear. Many servers only expose tools for platforms you have connected. In MegaMCP, platform tools appear once at least one account for that platform is switched on in Connections. Start a new chat afterwards so Claude reloads the tool list.
• It connected to the wrong workspace. Revoke the app in the provider's dashboard and connect again, choosing the right workspace at sign-in.
• A tool call fails with a permissions error. The account you signed in with may not have access to that ad account or property. Fix access in the platform, then refresh accounts.
• Claude answers without calling a tool. Name the platform or the account in your question, or say “use MegaMCP”. Models sometimes answer from memory when the question is vague.

Example prompts

Copy these into Claude, ChatGPT or Cursor with MegaMCP connected. Each lists the tools it typically calls; more are in the prompt library.

• Reads data

  List the integrations connected to this workspace and the tools each one adds.

  Tools it calls: megamcp_list_integrations

• Reads data

  Show me the demo account's campaigns that spent the most last month with the fewest conversions.

  Tools it calls: demo_ads_list_campaigns

• Proposes a change

  Pause the worst of those campaigns in the demo account so I can see what an approval looks like.

  Tools it calls: demo_ads_set_campaign_status

  Any change becomes a plan you review and approve. Nothing moves until you do.