Skip to content
Back to Guides

How to add an MCP server to ChatGPT

The OpenAI logo, labeled ChatGPT, joined to the Interfere logo by a line marked MCP
LG

Levi Gans

·

AI agents

ChatGPT can use tools from a remote MCP server in a conversation. You add the server on the web: open Plugins, select the plus button, choose Add custom MCP server, paste the server’s URL, set up authentication and select Create as a plugin. Install the plugin, then type @ in a chat and pick it.

The catch is who’s allowed to. Full MCP support is for ChatGPT Business and Enterprise/Edu, and in a Business workspace only admins and owners can add a server. Here’s the setup, then Interfere as a worked example.

Which ChatGPT plans support MCP servers

PlanWho can add a custom MCP serverWhat it can do
BusinessAdmins and owners onlyRead and write tools
Enterprise and EduAdmins and owners, plus members an admin grants it to with role-based accessRead and write tools
ProYouRead and fetch only

OpenAI’s help center doesn’t list Plus or Free. It’s web only: you can’t connect a custom MCP server from the mobile apps.

How to add a custom MCP server to ChatGPT

Older guides say to turn on developer mode first. OpenAI says the current Plugins experience doesn’t need it: you acknowledge a risk warning for each server instead.

  1. Go to ChatGPT Plugins on the web.
  2. Select the plus button, then Add custom MCP server.
  3. Enter a name and, optionally, a description.
  4. Under Connection, enter the server’s public HTTPS URL, including the path, usually /mcp.
  5. Configure authentication: OAuth, No authentication, or OAuth or no authentication.
  6. Review the risk warning and select I understand and want to continue.
  7. Select Create as a plugin and review the tools ChatGPT discovered.
  8. Install the plugin, start a new conversation, type @ and select it.

ChatGPT only connects to remote servers. For one on your laptop or a private network, OpenAI points you to Secure MCP Tunnel under Connection. Selecting a plugin applies to one message, so @mention it again for follow-ups.

What ChatGPT needs from an OAuth server

ChatGPT can’t send a custom API key, and it doesn’t support machine-to-machine grants like client credentials. OpenAI’s checklist:

  • Protected resource metadata, or a WWW-Authenticate header on 401 that points to it.
  • Authorization server metadata at /.well-known/oauth-authorization-server or /.well-known/openid-configuration.
  • PKCE with S256 in code_challenge_methods_supported. Without it, the server is unsupported.
  • A way to identify ChatGPT: a Client ID Metadata Document (CIMD), dynamic client registration (DCR) or a fixed client ID. ChatGPT prefers CIMD.
  • offline_access, so ChatGPT gets refresh tokens and you don’t sign in again.
  • Every advertised OIDC scope enabled for ChatGPT’s client. If the server advertises openid, email or profile, ChatGPT asks for them by default.

The last point bit us. Our identity provider advertised openid but didn’t allow ChatGPT’s client to request it, so the connection failed with invalid_scope until we allowed it. For the redirect URI, copy the exact one shown on the server’s management page in ChatGPT.

Read actions, write actions and confirmations

ChatGPT decides whether a tool reads or writes from the MCP readOnlyHint annotation. A tool without it counts as a write action, and write actions ask for confirmation by default. In Business and Enterprise workspaces, admins can change that under Workspace settings > General > Plugin permissions, for example to Allow read actions, which reads without asking and asks before changes.

Example: connect Interfere to ChatGPT

Interfere groups production errors into prioritized problems and explains each one with stack traces, session replays and affected users. Its MCP server has two URLs: https://mcp.interfere.com/mcp/readonly, where every tool is marked read-only, and https://mcp.interfere.com/mcp, which also allows writes.

Use the read-only URL. Interfere’s main tool, execute, runs any API operation, so on /mcp ChatGPT asks you to confirm every call, even reads. On /mcp/readonly, the server rejects writes and marks execute read-only, so Allow read actions covers it. If ChatGPT won’t finish sign-in on the read-only URL, add https://mcp.interfere.com/mcp instead and approve each call.

Add it

  1. In Plugins, select the plus button, then Add custom MCP server.
  2. Name it Interfere and enter https://mcp.interfere.com/mcp/readonly.
  3. Choose OAuth and leave static client credentials empty. Interfere supports CIMD and DCR, so ChatGPT registers itself.
  4. Accept the warning and select Create as a plugin. When ChatGPT asks, sign in to Interfere and pick a workspace.
  5. Review the discovered tools and install the plugin.

ChatGPT should find four tools: search finds API operations, execute runs them, and two more search Interfere’s docs. The connection is bound to the workspace you picked.

Ask about a production problem

For a first run, pick a problem you already understand:

Plain Text
1@Interfere List the Active problems in my production environment, highest2priority first. For the top one, explain the likely cause, the stack frame3it points to, how many users it affected, and which release recent4occurrences came from. Link the problem and two affected sessions.5Only use the Interfere app. Don't browse the web.

The last two lines follow OpenAI’s advice to name the app, or ChatGPT may reach for web search. Then check the evidence in Interfere. A deploy near the first occurrence is context, not proof.

Where ChatGPT stops

ChatGPT can read the problem but can’t fix it in your repository. OpenAI says its GitHub app “only lets you read from your repositories,” and points to Codex to edit and push code. To write the fix, take the problem link to Codex or Claude Code, which read the same Interfere server and work in your code.

Troubleshooting

What happensWhat to do
There’s no Add custom MCP server optionCheck your plan and role. On Enterprise, an admin can grant Create plugins with MCPs. Use the web, not the mobile app.
Sign-in fails with invalid_scopeEnable every scope the authorization server advertises, including openid and email, for ChatGPT’s client.
Sign-in fails with a redirect errorCopy the exact redirect URI from the server’s page in ChatGPT into the provider’s allowlist.
ChatGPT asks you to sign in again after a whileThe provider isn’t issuing refresh tokens. Advertise and enable offline_access, then recreate the plugin.
Interfere asks you to confirm every callYou’re on /mcp. Switch to /mcp/readonly.
Sign-in to /mcp/readonly doesn’t completeAdd https://mcp.interfere.com/mcp instead, and keep write confirmations on.

Where Interfere fits

We make Interfere, so weigh this part accordingly.

Coding agents now write code faster than teams can review it. Knowing what broke in production, who it hit and whether the fix held is the slow part. We built Interfere for that: it turns errors, traces and session replays into a problem with an explanation, and after you deploy a fix it watches until occurrences stop. ChatGPT is a good place to ask about that problem, especially for people who don’t live in a terminal.

Its SDKs cover Next.js, React with Vite and TanStack Start, and it takes OpenTelemetry from NestJS, Node.js and other services. There’s no mobile SDK, profiling, uptime or cron monitoring, or self-hosting. If you need those, another tool is the better fit.

Set up Interfere, connect it to ChatGPT, and ask it about one problem you already understand.