Connecting MCP to ChatGPT means adding a remote MCP server as a connector, then letting ChatGPT call that server's tools in chat. ChatGPT supports remote MCP servers, and OpenAI has published guidance for building them so they work with ChatGPT and Codex (OpenAI).
The pattern is spreading fast enough that API surveys now track it: Postman's 2025 State of the API, spanning more than 5,700 developers, architects, and executives, called out MCP as early but gaining momentum (Postman).
What You Need Before You Start
- A ChatGPT plan and account with connector access.
- A remote MCP server URL.
- A credential for that server — an API key or an OAuth grant, depending on the server.
MCP servers are either local or remote. ChatGPT connects to remote servers over HTTP. Cloudflare's overview is useful here: remote MCP servers are Internet-accessible and are what make MCP usable from web-based and mobile clients rather than only desktop developer tools (Cloudflare).
How Do You Add an MCP Server to ChatGPT?
- Open ChatGPT settings and enable developer mode (required for custom connectors).
- Go to the connectors/plugins area and choose to add a custom connector.
- Enter the server URL and give it a name.
- Choose the authentication method the server expects — API key or OAuth — and supply the credential.
Once saved, ChatGPT discovers the server's tools automatically. OpenAI notes that when you connect through a plugin, users in the workspace get an OAuth flow to the service if the server requires it.
How Do You Know It Worked?
Ask ChatGPT to do something read-only first — for example, "List the items on my server." A capability that returns data confirms three things at once: the URL is correct, the credential is valid, and the tool executed.
Only after a read-only call succeeds should you let the agent take actions. Write actions can have consequences, and OpenAI recommends keeping approval enabled for tools that modify data or take consequential steps.
What Are the Common Connection Errors?
| Symptom | Likely cause |
|---|---|
| "Couldn't reach server" | Wrong URL, or a rejected origin |
| Auth prompt loops | Credential wrong, expired, or revoked |
| Connected, no tools | Server exposes no tools, or the wrong scope |
| Tool fails mid-call | Server-side error or rate limit |
If the credential is an API key, treat a failure as a signal to rotate it. If the server rejects the client origin, that is a server-side allowlist issue, not a key issue.
Is Connecting ChatGPT to MCP Safe?
It depends entirely on what the server can do. OpenAI's own documentation warns that custom MCP servers are third-party services, and it walks through prompt-injection risks: malicious content in data an agent can read may attempt to make the model take unintended actions.
The caution is warranted. In Stack Overflow's 2025 survey of developers, 46% actively distrust the accuracy of AI tools while only 33% trust it, and 75% said they would still ask a human when they do not trust an AI answer (Stack Overflow). Practically: connect to servers you trust, prefer official servers from service providers, and keep write actions behind approval.
How Conbersa Works With ChatGPT
The Conbersa MCP server is a remote MCP server with API-key auth, so it connects to ChatGPT through the custom-connector flow. Add https://api.conbersa.ai/mcp with your cb_live_ key, then ask ChatGPT to list your devices. From there it can post to multiple accounts, run engagement, and schedule distribution. Full steps in Connect ChatGPT to Conbersa.