Guides
How to connect an MCP server to ChatGPT, step by step
Connect an MCP server to ChatGPT: plans that support it, the Plugins steps, tunnels for local servers, safety, and fixes when it fails.
6 min read

Quick answer
How do I connect an MCP server to ChatGPT?
Open Plugins in ChatGPT on the web and choose Add custom MCP server. Enter a name and the server's public HTTPS endpoint (or a Secure MCP Tunnel), set authentication, and accept the risk warning. Then select Create as a plugin. Full MCP with write actions needs Business or Enterprise/Edu; Pro gets read and fetch only.
Before you start: check three things
Most failed attempts to connect an MCP server to ChatGPT fail before the first click. Check these first.
1. Your plan
From OpenAI's help article on connecting custom MCP servers, as of November 2026:
- Business: only admins and owners can create and publish plugins for custom MCP servers in the workspace.
- Enterprise/Edu: admins and owners can, and they can use role-based access control to let selected members create connections too. Only admins and owners publish.
- Pro: you can build apps with the Apps SDK and connect custom servers with read and fetch permissions only. Tools that change things, such as generating a model or editing a scene, need full MCP.
- Mobile: not supported. Custom MCP connections are web only.
2. Where the server lives
ChatGPT connects to remote MCP servers over HTTPS, from OpenAI's side. It can't reach a server running only on your laptop. If yours is local (Blender MCP and Roblox Studio MCP both are), you need OpenAI's Secure MCP Tunnel, which connects a private server without exposing it to the public internet. Hosted services, Vi3W's MCP server among them, give you an HTTPS URL, so you can skip this.
3. Authentication
Find out whether the server uses OAuth, a token or nothing. If it's OAuth, OpenAI warns that the provider has to issue refresh tokens, usually through the offline_access scope. Otherwise ChatGPT loses access when the first login expires and you have to sign in again.
Step by step: add the server
In the current Plugins experience you don't turn on Developer Mode first.
- Open ChatGPT on the web and go to Plugins.
- Select the plus button, then Add custom MCP server.
- Enter a name and, optionally, a description. Pick a name you'll recognise after typing @.
- Under Connection, paste the server's public HTTPS MCP endpoint exactly as its documentation gives it, path included. For a local server, select Tunnel and pick your Secure MCP Tunnel.
- Configure authentication.
- Read the risk warning and tick I understand and want to continue.
- Select Create as a plugin, then review the tools ChatGPT discovered. This list is your best check that you connected the right server. A 3D server should list things like generate or export, not someone's CRM.
Use it in a chat
- Start a new conversation and type @, then select the plugin.
- Ask for something small and read-only first: "What can you do?" or "List my recent models."
- Then try a write action. ChatGPT asks for confirmation before write or modify actions, depending on the app's permissions and how risky the action is. Some especially risky actions are blocked outright.
Two habits save confusion. App selection applies to one message, not the whole chat, so @mention the app again when a follow-up needs it to act. You don't need to reselect it just to talk about results already in the conversation. And you can combine several apps in one prompt, for example fetching a reference image from one and generating a model with another.

Example: a 3D generation server
Here's what a sensible first session with a 3D MCP server looks like, using prompts that test one tool at a time:
- "@Vi3W make a low-poly wooden crate with iron corners, game-ready."
- "Show me what you made: polygon count, materials, file formats."
- "Add bones to it." (On a crate, expect the tool to refuse or produce something odd. A good test of how the server reports errors.)
- "Make a stylised fox, then add bones with the animal option and make it run."
- "Give me the GLB link."
Notice the pattern: name the app, ask for one thing, check the result, then build on it. For what Vi3W's server covers and how to add it, see ChatGPT apps for 3D models.
For workspace admins: publishing and updates
If you're setting this up for a team:
- Publish: after testing, admins and owners go to the Admin Console, select the workspace, open Plugins, choose Manage Legacy Apps, then Drafts and Publish, and review the safety warnings, especially for write actions. Enterprise/Edu admins can untick individual actions under Configure Actions and limit access to groups under Configure Access.
- Frozen tools: once approved, ChatGPT uses a frozen snapshot of the server's tools. When the server adds or changes a tool, nothing updates until an admin selects Refresh. New actions arrive switched off, and changed ones show as a diff. A breaking change can make calls fail until then, and users get no prompt telling them why.
- Deep research can use custom apps for read and fetch, not write.
Safety
OpenAI is blunt about it: connecting untrusted MCP servers raises your exposure to risks including prompt injection. Before you connect a server:
- Prefer servers whose publisher you can identify and whose documentation lists every tool.
- Be wary of servers that ask for broad OAuth scopes they don't need.
- For anything that edits files, test in a throwaway project first.
- Read is it safe to give an AI agent access to your 3D files? before pointing an agent at a real project.
What ChatGPT's MCP support can't do (yet)
- No local servers without a tunnel. Claude Desktop and Codex can launch a local server from a config file. ChatGPT on the web can't.
- No mobile. Set it up and use it on the web.
- No write actions on Pro. Read and fetch only.
- No automatic tool updates in workspaces. Admins refresh by hand.
If any of those blocks you, the same server usually works in Claude. Claude's custom connectors cover Free through Enterprise plans, and Claude Desktop also runs local servers. Background on how all this fits together: MCP explained for 3D artists.
The same server in Claude, for comparison
If you also use Claude, adding the same remote server takes a minute. On Pro or Max, go to Customize ‣ Connectors, choose + Add ‣ Add custom connector, give it a name, paste the server URL exactly as documented, review the authentication settings and pick a sign-in option, then Add. On Team and Enterprise, an Owner adds it and members click Connect. Claude reaches remote servers from Anthropic's cloud, so the same public-URL rule applies. Claude Desktop can also launch local servers from its config file, which ChatGPT on the web can't. Steps from Claude's help centre.
Connect Vi3W to ChatGPT
Vi3W's MCP server brings text-to-3D and image-to-3D, rigging, animation, retexturing and retopology into ChatGPT and Claude. It's hosted, so there's no tunnel to set up. Add it as a custom MCP server and @mention it. Try Vi3W.
Frequently asked questions
how to connect chatgpt to mcp server
On ChatGPT web, go to Plugins, press the plus button and choose Add custom MCP server. Give it a name, paste the HTTPS endpoint, configure authentication, tick the risk acknowledgement and select Create as a plugin. Then type @ in a chat and pick it.
chatgpt mcp without developer mode
In the current Plugins experience you don't need Developer Mode. You add the server from Plugins and acknowledge a risk warning for each server instead. Older guides that start with "enable Developer Mode" describe the earlier flow.
chatgpt mcp free plan
OpenAI's help article lists custom MCP connections for Business and Enterprise/Edu (full MCP) and Pro (read and fetch only). It doesn't list the Free plan. If you need MCP on a personal budget, Claude's custom connectors are available on its Free plan.
chatgpt mcp not working
Check four things. You're on the web, not mobile. The endpoint is public HTTPS, or a tunnel for a local server. Your plan allows the action (Pro is read-only). And, in a workspace, the admin has refreshed the tool list after the server changed, because ChatGPT uses a frozen snapshot of the tools.
Last reviewed

