ChatGPT connectors and apps for marketing

Connectors, apps and MCP: the naming, untangled

OpenAI has used several names for the same idea. Early integrations with Google Drive, SharePoint and similar services were called connectors. As ChatGPT grew a directory of third-party integrations that can also show interactive interfaces, the name became apps. Many help articles, and most people, still say connectors.

The technology underneath is the Model Context Protocol. An app is backed by an MCP server that lists tools; ChatGPT reads the tool descriptions and calls them when a question needs them. That is why one MCP server can serve ChatGPT, Claude, Cursor and VS Code at once: they all speak the same protocol.

What marketers actually use them for

The useful pattern is simple: questions that today need an export, a spreadsheet and twenty minutes become one sentence, and fixes that today need clicking through an ad platform become a reviewed plan.

• Weekly reporting. Pull spend, conversions, CPA and ROAS for every account and have ChatGPT write the summary your client or boss reads.
• Wasted-spend hunts. Read the search terms report, find queries that spent without converting, and propose negative keywords.
• Analytics questions. “Why did key events drop?” answered from GA4 with the breakdown, not a guess.
• SEO chores. Find striking-distance queries in Search Console, then draft better titles and meta descriptions for the pages that rank for them.
• CRM loops. See which campaigns produced deals in HubSpot, not just form fills.

What they are not good for, yet: anything where a wrong action is expensive and nobody checks. A model can misread “pause the bad ones” as all of them. The fix is structural, not a better prompt, and it is covered below.

How to add a custom marketing app to ChatGPT

Custom apps need developer mode, which OpenAI offers on paid plans (Plus, Pro, Business, Enterprise and Edu at the time of writing). On business plans an admin may have to allow it first.

1. Turn on developer mode.

   On chatgpt.com, open Settings, find developer mode (under the security or apps settings) and switch it on.

2. Create an app.

   In ChatGPT's apps (connectors) settings, choose to create or add an app.

3. Paste the server URL.

   For MegaMCP that is https://megamcp.com/api/mcp. Name it MegaMCP and choose OAuth as the authentication method.

4. Sign in.

   ChatGPT opens the provider's sign-in page. With MegaMCP you sign in, pick your workspace and approve access. New workspaces get a 5-day trial and a sandbox ad account.

5. Use it in a chat.

   Start a new chat, pick the app from the tools menu, and ask a read-only question first.

The current, maintained version of these steps is at Connect ChatGPT. Menu names move, so if something looks different, trust OpenAI's help center over any screenshot.

Read tools vs write tools, and why it matters

MCP lets a server mark each tool with hints, including whether it only reads. ChatGPT uses those hints to decide how cautious to be. A well-built marketing app marks every report as read-only and makes its write tools explicit.

MegaMCP goes one step further: write tools never touch the platform. A write tool reads the live state and returns a plan, one line per change with before and after values. The plan waits on the Changes page of the dashboard until a person approves it. Approving in chat is possible, but it is off by default and an admin has to opt in, and even then only for plans the same connection proposed.

Three more layers sit underneath: budget bumpers (per-account spend caps the AI cannot edit), undo for supported changes, and a flight recorder of every tool call, approval and reversal.

ChatGPT or Claude for marketing work?

Both support remote MCP servers, so you don't have to choose your tools by AI app. Teams usually pick based on what they already pay for and how they like to work:

• ChatGPT suits teams already on ChatGPT Business or Enterprise, and people who like its app directory and interactive app views.
• Claude suits long, careful analysis and writing, and teams who want an organization-wide connector that each member signs into. See Claude connectors for marketers.

Because the server is the same, you can connect both and compare answers on the same question. That is one of the better ways to catch a model's mistake.

Limits worth knowing before you rely on it

• Tool lists are long. Marketing servers can expose dozens of tools. If ChatGPT picks the wrong one, name the platform and the account in your prompt.
• Data freshness is the platform's. GA4, for example, can take a day or more to finish processing recent data. An app can't make data fresher than the source.
• Context windows are finite. Ask for summaries and top-N lists rather than every row; good tools return compact tables for exactly this reason.
• Approval reviews take time. Directory listings and some platform connections depend on reviews by OpenAI and by the ad platforms themselves.

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

  Give me a weekly summary of every ad account I can paste into Slack: spend, conversions, CPA, ROAS and the one thing to look at.

  Tools it calls: megamcp_cross_channel_summary

• Reads data

  List Search Console queries where we rank 8 to 20 with at least 500 impressions.

  Tools it calls: gsc_quick_wins

• Reads data

  Which HubSpot deals closed this quarter, and what was each contact's original traffic source?

  Tools it calls: hubspot_searchhubspot_get_record

• Proposes a change

  Add the junk search terms from the demo account as negatives and show me the plan.

  Tools it calls: demo_ads_search_termsdemo_ads_add_negative_keywords

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