Last updated: November 2026
To add a custom MCP server to ChatGPT, open ChatGPT on the web, go to Plugins, select the plus button and choose Add custom MCP server, then enter the server URL and authentication, accept the risk warning and select Create as a plugin. At the time of writing, OpenAI's help center lists full ChatGPT MCP support, including write actions, for Business, Enterprise and Edu, and custom servers with read and fetch tools only for Pro; Plus and Free are not listed. The server must be a remote MCP server reachable over public HTTPS (or through OpenAI's Secure MCP Tunnel), not a local stdio server running on your computer.
This guide covers the steps for any custom MCP server, then walks through a real example with Automata, a free, open-source MCP server for WhatsApp, Telegram, email and PostgreSQL.
What is an MCP server, and what can ChatGPT do with one?
MCP (the Model Context Protocol) is an open standard that lets an AI app call tools that live somewhere else. An MCP server is the program that offers those tools, such as "search my messages" or "run a SQL query". It describes each tool, and ChatGPT decides when to call one.
Once you add a server, ChatGPT can:
- Read data it can't otherwise see, such as your database, inbox or chat history.
- Act, if the server offers write tools and your plan allows them: send a message, update a record, start a workflow.
- Combine tools from several servers in one request.
ChatGPT only sees what the server's tools return, so the server is where you decide what the AI can reach.
Before you start
1. A plan that supports custom MCP servers, used on the web. OpenAI's help article "Connecting custom MCP servers in ChatGPT" says, at the time of writing:
| Plan | Can add custom MCP servers | Write tools |
|---|---|---|
| Business | Yes, admins and owners | Yes, with confirmation |
| Enterprise, Edu | Admins, owners and members granted permission through role-based access | Yes, with confirmation |
| Pro | Yes | No, read and fetch only |
| Plus, Free | Not listed | Not listed |
Custom MCP servers are web only; you cannot add one in the ChatGPT mobile apps. On Business, Enterprise and Edu, the workspace permission is called Create plugins with MCPs.
2. A remote MCP server URL. ChatGPT connects from OpenAI's servers, so localhost never works. You need a public HTTPS endpoint that speaks streamable HTTP (usually ending in /mcp) or SSE, or a private server connected through Secure MCP Tunnel.
3. To know whether your tools read or write. ChatGPT checks each tool's readOnlyHint annotation. Tools marked read-only run without asking. Any tool without the hint counts as a write action and asks for your confirmation first. On Pro, only read and fetch tools are available.
What about developer mode? In 2025, adding a custom connector with write tools meant switching on a beta setting called developer mode. In the current Plugins experience, OpenAI says you no longer need to turn on developer mode: you add the server from Plugins and acknowledge a risk warning for each one. Whether write tools work now depends on your plan and workspace permissions.
How to add a custom MCP server in ChatGPT
These steps follow OpenAI's help center and developer docs as checked in November 2026. ChatGPT's menus changed several times this year, so labels in your account may differ slightly.
Step 1: Open Plugins and add a custom MCP server
- Open ChatGPT Plugins on the web.
- Select the plus button, then Add custom MCP server.
- Enter a name and, optionally, a description. The description helps ChatGPT decide when to use the server, so say what it is for: "Read-only access to our sales tables."
Step 2: Enter the server URL and authentication
- Under Connection, enter the server's public HTTPS URL, including its path (often
/mcp). For a private server, select Tunnel and choose a Secure MCP Tunnel or enter itstunnel_id. - Choose the authentication method:
- OAuth: you sign in to the server once, and ChatGPT attaches the access token to every request. Best for servers that hold personal or company data.
- No authentication: ChatGPT sends no credentials. Use it for public data, or for servers that put an access token in the URL itself.
- OAuth or no authentication: some tools work anonymously and others ask you to sign in.
ChatGPT cannot send a custom API key or Authorization header. If your server only accepts a bearer key, check whether it offers OAuth or a token-in-URL endpoint.
Step 3: Create, review and install the plugin
- Read the risk warning and select I understand and want to continue.
- Select Create as a plugin.
- Check the tools ChatGPT discovered. This is your chance to spot a write tool you did not expect.
- Install the plugin. It appears in your personal plugins, or in the workspace where you created it.
Each plugin has a details page in settings, where you can turn individual tools off and refresh the tool list after the server changes.
Step 4: Use it in a chat and approve tool calls
- Start a new conversation, type
@in the prompt box and select the plugin. The selection applies to that message only, so@-mention it again when a follow-up needs fresh data. - When ChatGPT wants to call a write tool, it shows a confirmation with the exact input. Read it before you approve. You can remember the decision for that conversation; new conversations ask again.
- Expand any tool call to see its full JSON input and output.
Worked example: connect Automata to ChatGPT
Automata is an MIT-licensed MCP server, hosted or self-hosted, for WhatsApp, Telegram, email (IMAP, read-only plus drafts) and PostgreSQL. Each connector token is bound to one connection and limited to the actions, chats, folders or tables you choose.
1. Sign up at mcp.getautomata.app/signup. The hosted free plan covers one connection and one connector token; Pro is $19/month. A self-hosted instance also works if ChatGPT can reach it over public HTTPS.
2. Add a connection. Pick one:
- WhatsApp: scan the QR code from Linked devices on your phone.
- Telegram: choose Show QR code and scan it from Telegram → Settings → Devices → Link Desktop Device.
- PostgreSQL: paste the connection string of a least-privilege, read-only role, ideally on a replica.
3. Create a scoped connector token. Select New connector token and set:
- Name:
chatgpt, so you know what to revoke later. - Expires: 30 or 90 days.
- Chats or Tables: choose Only selected chats or Only selected tables, and pick only what ChatGPT needs.
- Actions: start read-only. WhatsApp and Telegram tokens include the send tool by default, so choose Only selected actions and untick it. PostgreSQL and email tokens start read-only.
Select Create token and copy it. Automata shows it once and stores only a hash.
4. Paste the connector URL into ChatGPT. Because ChatGPT cannot send an Authorization header, use the path-token URL, where the token is part of the address:
https://mcp.getautomata.app/mcp/<connector-token>In Add custom MCP server, paste it as the server URL and set authentication to No authentication; Automata checks the token on every request. A read-only token shows four tools, and Automata marks each with readOnlyHint, so ChatGPT runs them without asking.
Anyone with the full URL gets the access the token allows. Keep it out of screenshots and shared chats, and revoke the token in Automata if it leaks.
5. Test it. In a new chat, @-mention the plugin and try:
- "Summarize what's new in the Project X group since Monday." Automata's WhatsApp and Telegram tools don't track read status, so ask by date rather than for "unread" messages.
- "What were last month's top 10 orders?" (PostgreSQL)
- Then ask about a chat or table you did not share. Automata refuses it, and ChatGPT should say so.
Local MCP servers and ChatGPT
Many open-source MCP servers on GitHub, including most "run it with npx" servers, use the stdio transport: a desktop app such as Claude Desktop or Cursor starts them as a local process. ChatGPT can't do that. It runs in OpenAI's cloud and can only call a URL.
Your options:
- Host it remotely. Deploy the server with HTTPS and streamable HTTP. A stdio-only server can sit behind a bridge such as
mcp-proxyor Supergateway. You then own authentication: never publish an unauthenticated server that reads private data. - Use a tunnel. OpenAI's Secure MCP Tunnel runs a
tunnel-clientinside your network that only makes outbound connections and can reach a stdio or HTTP server. It needs a tunnel ID from OpenAI's platform settings and an API key. For quick tests, ngrok or Cloudflare Tunnel also give you an HTTPS URL. - Use a hosted server. It already has an HTTPS endpoint and authentication, so there's nothing to deploy.
Keeping it safe
- Give each connector the least access it needs. Scope by chat, folder, table or tool on the server side. Model instructions are not a security boundary; server-side limits are.
- Start read-only. Add write tools later, on a separate connector with a narrower scope, once you trust the setup.
- Review every write approval. Read the tool input before you confirm, and only remember approvals for servers you fully trust.
- Expect prompt injection from returned data. A message, email or database row can contain instructions aimed at the AI, and reading it is enough to put them in front of the model. Avoid pairing a server that reads untrusted content with one that can send data out in the same chat.
- Remember where data goes. Everything a tool returns becomes part of your ChatGPT conversation and is sent to OpenAI.
- Know how to revoke access. Remove or disconnect the plugin in ChatGPT, and revoke the credential on the server side: an OAuth grant, or with Automata, the connector token. Use one credential per client so revoking ChatGPT doesn't break anything else.
Troubleshooting
| Problem | Likely cause and fix |
|---|---|
| Add custom MCP server is missing | Your plan isn't eligible, you're on mobile, or a workspace admin hasn't given you Create plugins with MCPs. |
| Plugin doesn't show in the chat | Install it after creating it, then type @ and pick it for this message. |
| Server not reachable | The URL isn't public HTTPS, has the wrong path, or the server doesn't speak streamable HTTP or SSE. localhost never works. Test it with npx @modelcontextprotocol/inspector. |
401 or authentication errors | The token or OAuth grant is wrong, expired or revoked. For OAuth, the provider must issue refresh tokens (offline_access) or you'll be signed out when the first token expires. |
| ChatGPT answers without calling tools | Name the plugin and tool: "Use Automata's run-query tool. Don't answer from memory." |
| Write tools or developer mode missing | Pro is read and fetch only, and the current Plugins experience has no developer mode to turn on. Writes need Business, Enterprise or Edu. |
| New tools don't appear | Refresh the plugin. In a workspace, an admin must review and publish tool changes. |
What you can connect
| Service | What ChatGPT can do | Guide |
|---|---|---|
| Read, search and summarize the chats you allow; send only if the token allows it | ChatGPT + WhatsApp guide | |
| PostgreSQL | Read-only SQL on the tables you choose | ChatGPT + PostgreSQL guide |
| Telegram | Read and search chats, groups and channels; send only if allowed | Telegram MCP server |
| Email (IMAP) | Read and search allowed folders; save drafts, never send | Email connection docs |
Frequently asked questions
Does ChatGPT support MCP?
Yes. ChatGPT supports remote MCP servers as custom plugins on the web. You add one from Plugins with Add custom MCP server, and ChatGPT can call its tools in conversations. It supports the streamable HTTP and SSE transports, with OAuth or no authentication.
Which ChatGPT plans support custom MCP servers?
At the time of writing, OpenAI's help center lists full MCP support, including write actions, for Business, Enterprise and Edu, and read and fetch tools only for Pro. Plus and Free are not listed. On Business, only admins and owners can add servers; Enterprise and Edu admins can grant the permission to other members.
Can ChatGPT use a local MCP server?
Not directly. ChatGPT connects from OpenAI's servers, so it can't start a local stdio server or reach localhost. Host the server at a public HTTPS URL, connect it through OpenAI's Secure MCP Tunnel, or use a hosted MCP server instead.
What is ChatGPT developer mode?
Developer mode was the beta setting, introduced in 2025, that let you add custom MCP connectors with write tools. In the current Plugins experience, OpenAI says you no longer need to turn it on: you add a custom MCP server from Plugins and accept a risk warning for each one. Your plan decides whether write tools are available.
Is it safe to connect ChatGPT to my apps?
It can be, if access is limited on the server side. Give each connector only the data and tools it needs, start read-only, read every write confirmation, and connect only servers you trust. Returned data can carry prompt injection, and all of it is sent to OpenAI.
Can ChatGPT send messages through an MCP server?
Yes, if the server has a send tool and your plan allows write actions (Business, Enterprise or Edu). ChatGPT asks for confirmation before each write by default. With Automata, sending is a per-token action: leave it off and the send tool isn't registered at all.
Sources
- OpenAI Help Center, Connecting custom MCP servers in ChatGPT, checked November 16, 2026.
- OpenAI developer docs, Connect a custom MCP server, Connect and test your plugin, Authentication and Secure MCP Tunnel.
- Automata docs: MCP clients and Tokens and permissions.
Comments
Questions, corrections, or your own setup? Join in below. Pick any name to post.