ChatGPT MCP: How to Add a Custom MCP Server (October 2026)

Does ChatGPT Support MCP?

Yes. As of October 2026, ChatGPT connects to MCP servers through plugins: open Plugins on chatgpt.com, select the plus button, choose Add custom MCP server, enter a remote HTTPS URL and its authentication, then select Create as a plugin and install it. ChatGPT supports both read and write tools from your server, but it only connects to remote servers (SSE or streamable HTTP) — a local stdio server needs a tunnel, or the Codex side of the ChatGPT desktop app.

The names have changed several times in eighteen months — connectors, then apps and "developer mode", now plugins — so many guides that rank for "chatgpt mcp" describe menus that no longer exist. This guide follows OpenAI's current Add custom MCP server and plugin quickstart docs.

How ChatGPT's MCP Support Evolved

Two things stayed constant: ChatGPT talks to MCP servers over the network, and OpenAI does not review the custom servers you add. Today a plugin can bundle an MCP server, skills (reusable instructions) or both. An MCP server you add yourself is turned into a plugin that only you — or your workspace — can see.

Who Can Add a Custom MCP Server?

OpenAI's developer docs say adding a custom server happens in ChatGPT on the web, and that workspace permissions and security restrictions, including Lockdown, apply. Plan rules are the part that changes most often, so treat this table as a starting point and check your own Plugins page:

In a Business, Enterprise or Edu workspace, an admin can also limit a connection to read-only actions or an approved set of actions, and decide when ChatGPT must ask before using it. If you do not see Add custom MCP server, the usual cause is a workspace policy rather than a bug.

Step-by-Step: Add a Remote MCP Server to ChatGPT

You need an MCP server reachable at a public HTTPS URL, usually ending in /mcp. If you just want to try the flow, OpenAI's quickstart uses a public demo server with a read-only roll_dice tool and no authentication: https://tinymcp.dev/api/moldy-aloof-zettabyte/mcp.

1. Go to Plugins in ChatGPT on the web.
2. Select the plus button, then Add custom MCP server.
3. Enter a name and, optionally, a description.
4. Under Connection, paste the Server URL including the /mcp path. For a private server, select Tunnel instead (see below).
5. Choose Authentication: OAuth, No authentication, or "OAuth or no authentication" (mixed).
6. Read the risk warning and select I understand and want to continue.
7. Select Create as a plugin. ChatGPT connects and lists the tools it discovered.
8. Open your personal plugins, open the new plugin and select the plus button to install it.
9. Start a new chat, type @ in the prompt box and pick the plugin.

OpenAI's quickstart tests the plugin in Work mode (the Chat/Work switch at the top of the ChatGPT home page), but installed plugins can be used from both Chat and Work.

Authentication options

• No authentication — fine for public, read-only servers such as documentation search or DeepWiki.
• OAuth — for anything that touches a user's account. If you enter static client credentials, ChatGPT uses them; otherwise it can register itself with Client ID Metadata Documents (CIMD) or Dynamic Client Registration (DCR) when your authorization server supports them. Server authors should follow the MCP authorization spec.
• Mixed — initialize and tool listing run without auth, then each tool uses OAuth or no auth according to the security scheme in its metadata.

There is no field for a raw API key header in this flow. If a server only accepts a static key, put an OAuth layer or proxy in front of it, or use it from Codex instead, which supports bearer tokens.

Using MCP Tools in a Conversation

ChatGPT picks tools from their names and descriptions, so explicit prompts work best when several plugins overlap:

• Name the plugin and tool: Use the Acme CRM plugin's update_record tool to …
• Exclude alternatives: Do not use web search; only use Acme CRM.
• Give the order: First call read_file, then write_file with the edited content.

Write actions need approval. ChatGPT respects the MCP readOnlyHint annotation; any tool without it is treated as a write and asks for confirmation by default. You can expand each tool call to see the full JSON input and output before approving. Remembering "approve" applies only to the current conversation. Since June 2026 you can also choose when connected apps ask permission — always, before making changes, or only before important changes — and Business and Enterprise admins set the workspace defaults.

Refreshing tools. ChatGPT caches the tool list. After you change tool names, descriptions or schemas on the server, open the plugin's details page and select Refresh, then start a new conversation. The same page lets you turn individual tools on or off.

Connecting a Local or Private MCP Server

ChatGPT's servers cannot reach localhost or your office network. You have three options:

1. Secure MCP Tunnel — OpenAI's supported route. Create a tunnel in Platform settings, run tunnel-client inside your network (it only needs outbound HTTPS to api.openai.com:443), then pick Tunnel under Connection. It works with stdio or HTTP servers behind your firewall. See Secure MCP Tunnel.
2. A temporary HTTPS forwarding service — fine for development testing, not for anything you share.
3. Codex in the ChatGPT desktop app — Codex runs on your machine and launches local stdio servers like Filesystem or Playwright directly from ~/.codex/config.toml. Hosted ChatGPT chats on the web do not read that file. See our Codex MCP setup guide.

To bundle a whole stack of local servers behind one remote endpoint, a gateway like the one in our Docker MCP Gateway guide plus a tunnel is a common pattern.

Good Remote Servers to Try in ChatGPT

Before you add any of them, check whether an official plugin already exists in the Plugins directory. Official plugins are reviewed by OpenAI and usually bundle skills alongside the server.

Building Your Own: Apps SDK Is Now "Plugins"

If you are building rather than connecting, the former Apps SDK documentation now lives under Plugins. The core is still an MCP server over streamable HTTP. Since February 2026 ChatGPT is fully compatible with the open MCP Apps spec for optional UI, so a server can return interactive components that also work in other MCP Apps hosts. ChatGPT also reads the MCP instructions field at initialization — keep the first 512 characters self-contained.

For a quick test outside ChatGPT, use MCP Inspector to list and call tools, or the Responses API, which accepts any remote server as a tool:

curl https://api.openai.com/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "model": "gpt-6-astra",
    "tools": [{
      "type": "mcp",
      "server_label": "openai_docs",
      "server_url": "https://developers.openai.com/mcp",
      "require_approval": "never"
    }],
    "input": "Search the OpenAI docs for Responses API streaming."
  }'

Public distribution goes through the plugin submission portal and needs a stable public HTTPS endpoint. Plugins with a local stdio server are marked Desktop only, install only in the ChatGPT desktop app, and are currently limited to approved partners.

Limitations and Troubleshooting

• No "Add custom MCP server" button. Your plan or workspace policy does not allow it. Ask a workspace admin, or check whether you are in a personal or workspace account.
• "Couldn't connect" on create. Verify the URL with MCP Inspector, confirm the /mcp path, and make sure your proxy or CDN does not buffer streaming responses.
• No tools listed. The server is reachable but returned nothing from tools/list; check server logs and schema errors.
• 401 loops. Return a WWW-Authenticate header that points to your /.well-known/oauth-protected-resource metadata so ChatGPT knows to restart OAuth.
• New tools don't appear. Select Refresh on the plugin and start a new conversation.
• Wrong tool chosen. Rewrite tool descriptions with "Use this when…" guidance and split overlapping tools — see how MCP tool discovery works.
• Security. Custom servers are not verified by OpenAI. A malicious server can exfiltrate chat data or inject instructions, so read MCP security best practices and keep write approvals on.

Next Steps

Browse the MCP server directory for servers that offer a remote HTTPS endpoint, filter by category, and compare how other MCP clients connect to the same servers.