Without context, AI takes a guess: sometimes it lands well, sometimes it misses and that's frustrating.
With more context, the result can only be better!
You can try hard to put context into a prompt (that's a first step), but you can also rely on the AI's ability to go and fetch the information it needs to reach the given objective.
For that, models need to connect where our information is stored: knowledge base, database, everyday tools…
Add to that the ability to act and you get a real copilot.
And that's exactly the shift happening right now: an active AI. This is now possible thanks to ChatGPT connectors, the model has access to the data and to the actions, directly from the conversation.
Enabling connectors in ChatGPT
To connect ChatGPT to tools, it happens in Settings → Connectors.
Some tools can already be integrated in 1 click. To add others, you need to go through developer mode in the section's advanced settings.
At that point, you can add “MCP servers” (it may sound like gibberish, but it's not that complicated).
MCP, what is it in a few words?
The Model Context Protocol (MCP) standardizes how a model discovers tools, understands their parameters and calls them. It's a bit like the USB-C of AI models. On one side (that of the tools you use), you find:
- Tools : typed actions (such as create_invoice, search_pages) with schemas (required inputs, formats)
- Resources (optional) : readable files/knowledge sources
- Authentication : how to connect to the app (OAuth, key)
On the ChatGPT side, the model:
- Discovers the available tools at the start of the conversation.
- Chooses the relevant tool when your request calls for it.
- Fills in the parameters (and asks you for the missing ones).
- Executes the action (with or without confirmation depending on your settings).
Result: the model has more capabilities encapsulated that the model orchestrates on its own. Schematically, it looks like this:
Concretely, what can you do?
Here are 5 concrete cases to help you picture using connectors:
1️⃣ Canva: creating a branded infographic
Typical command: “Generate a 1080×1350 infographic about AI agents, palette #0F172A/#22D3EE, 2 variants.”
2️⃣ Notion: searching for information, creating a page, etc.
Typical command: “Create a Sprint 14 Retro page in Ops/Sprints, summarize the 10 ‘bug’ tickets from the last 7 days, fields: Owner, Impact, Next step.”
3️⃣ Fireflies → Meeting notes → CRM / Gmail
Typical command: “Pull yesterday's ACME discovery call, extract decisions + objections, push into Notion Sales/Meetings + prepare a draft email.”
4️⃣ Stripe: creating an invoice, a customer, and more
Typical command: “Create an invoice for ACME of 799 € excl. tax, due in 15 days, description AI Workshop, send the link to the main contact.”
Going further
With automation tools like Zapier, you can use their MCP server and benefit from thousands of applications connected to ChatGPT.
An example of integrating Zapier and ChatGPT:
A word about this edition's partner: Brevo 💚
Since we're talking marketing, I'll take the chance to tell you about Brevo : a complete CRM and mailing tool.
Very competitive on the mailing side (pay as you go), it avoids being overcharged as your contact base grows. And you'll no longer get emails like:
CRM management pairs with mailing, and you can set up drag-and-drop automations similar to Zapier.
You can try Brevo for free today or use the code LOUISGRAFFEUIL to get 50% off for 3 months (annual plan).
👉 Head over here if you're interested
Your own MCP server
You can go even further by creating your own MCP server via n8n. This lets you connect your own tools and local data, with the advantage of being on your own server.
To do this, you need to retrieve the “Production” URL of the MCP Server node and add it to ChatGPT, then expose the actions you need (database, CRM, calendar, ..)
List of ready-to-use servers
You can take a look at this link to get the list of MCP servers available in ChatGPT.
Safeguards to avoid security issues
Yes, there are attempts at prompt injection (for example with external content in a calendar/email to change the model's instructions). Worth mentioning, no need to panic. My 80/20:
- Minimal scopes : grant only the necessary permissions.
- Check sensitive actions (sending an email, an invoice) → ChatGPT always asks for confirmation
- Logging is possible (a Notion “AI Actions” page).
- Adjustment in the prompt: “if in doubt, ask for more info, do not execute”
