ChatGPT Developer Mode lets you connect any remote MCP server and call its tools from a chat. See who can enable it, where the toggle lives, how to fill the connector form, which errors appear first, and how to keep write actions safe in practice.
Paste a URL into ChatGPT, flip one switch, and the chatbot can suddenly read your tickets, query your database, or push a change to a repo. That is what ChatGPT Developer Mode MCP support does: it turns ChatGPT into a full Model Context Protocol client that can call your tools, not just read from them. The catch? The toggle has moved around during 2026, the rules change with your plan, and the first connection fails more often than anyone admits.
This article walks through the whole path in the order you will actually hit it. You will see who can turn the feature on, where the switch lives, how to fill in the connector form, what to do with a server that only runs on your laptop, which errors show up first, and which safety habits are worth the extra thirty seconds. The steps below come from OpenAI's help center and from setup notes published by teams that have already connected servers this way.
💡 Heads up: OpenAI has relocated the Developer Mode toggle more than once this year. Treat the menu paths below as landmarks, and if a label does not match your screen, search the settings panel for "Developer Mode".
Model Context Protocol, or MCP, is an open standard that lets an AI client call tools exposed by a server. Think of it as a USB port for software: any server that speaks the protocol can plug into any client that does. Out of the box, ChatGPT talks to a short list of vetted integrations. Developer Mode removes that limit. You paste the URL of any remote MCP server, and ChatGPT treats it as a first-class source of tools.
OpenAI now uses three words that people mix up constantly, so it helps to pin them down:
| Term | What it means | How you add it |
|---|---|---|
| Connectors | First-party integrations such as Google Drive and GitHub | Pick from the built-in list |
| Apps | Third-party servers that OpenAI reviewed and listed in a directory | Pick from the directory |
| Developer Mode | Any remote MCP server, including the one you built last night | Paste its URL yourself |
This article is about the third row. Listed apps skip the setup entirely, but they also skip your control over what the server does.
Older connectors were mostly about reading and searching. Developer Mode gives ChatGPT the full protocol, which means write tools: create a ticket, update a CRM row, trigger a deploy, or chain several servers in a single request. Before a write call goes out, ChatGPT shows a confirmation with the full JSON payload, so you can read exactly what is about to be sent. Tools that a server flags with readOnlyHint skip that prompt, which is why honest hints matter a lot (more on that in the safety section).
💡 Rule of thumb: if a tool changes anything outside the chat, assume it will ask for approval, and assume you should read that approval before clicking.
Developer Mode and custom MCP connectors belong to paid plans, and they live in the web app. The free tier does not get them.
| Plan | Developer Mode | Who flips the switch |
|---|---|---|
| Free | Not available | Nobody |
| Plus | Available | You |
| Pro | Available | You |
| Business | Available | A workspace admin first, then members |
| Enterprise and Edu | Available | A workspace admin first; authorized developers can then test servers privately |
On Business, Enterprise and Edu workspaces, an admin has to enable the feature before anyone sees it. The setting sits in workspace settings under Permissions & Roles, in the connected data section, with a label along the lines of "Developer mode / Create custom MCP". If you are a member and the Create button never shows up, that is the first thing to ask your admin about.
Here is the usual route:
Because the toggle has been moving, you may see it in a different spot:
| Path | Where people report it |
|---|---|
| Settings → Connectors → Developer Mode | Older tutorials and many setup notes |
| Settings → Apps & Connectors → Developer Mode | Newer builds |
| Settings → Security and login → Developer Mode | Recent write-ups |
Two things change once it is on. A Create button appears on the connectors page, and a Developer Mode option shows up in the + menu of the message box. If you can see both, you are ready.
Three requirements decide whether the connection can work at all, so check them before you open any form:
localhost URLs fail./mcp or /sse. Open it in a browser first. A live MCP endpoint answers with something, even if that something is an error about missing headers. A dead one just times out.Go to Settings → Connectors, press Create, and fill in the form:
| Field | What to enter |
|---|---|
| Name | A short label you will recognize in the + menu |
| Description | One line saying what the server does |
| Connector URL | The full public HTTPS address, for example https://mcp.example.com/mcp |
| Authentication | None, OAuth, or a token option if your build shows one |
| Trust checkbox | "I trust this provider": tick it only after you have read what the server exposes |
Press Create. ChatGPT contacts the server, pulls the tool list and displays it. Switch off any tool you do not plan to use. A smaller menu means fewer wrong guesses later.
This choice matters more than it looks, because it decides who else can call your tools.
| Option | Pick it when | Watch out for |
|---|---|---|
| None | The server genuinely has no auth and only exposes public, read-only data | Anyone holding the URL can call every tool |
| OAuth | The server supports sign-in, which is true of most vendor servers and anything touching private data | Login loops if you start sign-in in one browser and finish it in another |
| Token | Your build offers it and the server expects a bearer token or API credential | The credential is pasted once, so rotate it when a teammate leaves |
Choose the lowest-privilege option that still works, and never point None at anything you would not post on a public website.
Connecting a server is not the same as using it. The connector has to be enabled in each conversation:
Explicit wording beats a vague request every time:
Use the Acme Tickets connector to list my open tickets from this week.
Read only. Do not create or edit anything.
When a write call comes up, read the JSON payload in the confirmation before you approve it.
💡 Deep research is different. In that mode ChatGPT only uses two tools named
searchandfetch, and any other tool is ignored. A server that exposes neither will not work there, even though it works fine in normal chat.
ChatGPT cannot reach localhost, so a tunnel gives your laptop a temporary public HTTPS address. ngrok and cloudflared are the usual choices. A typical session looks like this:
# terminal 1: start your server
python server.py
# terminal 2: open a tunnel to its port
ngrok http 8000
# connector URL to paste into ChatGPT
https://abc123.ngrok.io/mcp/
Mind the path at the end. A server built with FastMCP exposes /mcp/, and a missing path or a stray slash is a classic reason for a failed first connection.
A tunnel opens a door from the internet to a port on your machine. Use throwaway data, expose only read-only tools while you test, keep real secrets out of the environment, and close the tunnel when you finish. Once the server behaves, move it to a stable host with proper authentication instead of living behind a temporary URL.
Most failures fall into a handful of patterns:
| Symptom | Likely cause | Fix |
|---|---|---|
| No Create button | Developer Mode is off, you are on the free plan, or your admin has not enabled it | Check the plan, then ask the workspace admin |
| "Connection failed" | The URL is not HTTPS, not public, or points at the wrong path | Open it in a browser and double-check the endpoint |
| 401 or 403 errors | Expired or wrong token, or an OAuth session that lapsed | Regenerate the credential or reconnect |
| No tools appear | Developer Mode is not enabled in this chat, or the tool list is stale | Start a new chat and refresh the connector |
| Tool never gets called | ChatGPT chose another source | Enable the connector from the + menu and name the tool in your prompt |
| Login loop | A stale OAuth session | Disconnect, reconnect, and finish sign-in in the same browser |
When ChatGPT says the connection failed but your browser opens the URL just fine, test with the MCP Inspector, a debugging client for MCP servers. If the Inspector cannot list your tools, ChatGPT will not either, and you have saved yourself a round trip.
Log every incoming request with its timestamp and path. A 401 in the log points at authentication. No log line at all means the request never arrived, so look at DNS, the tunnel, or a firewall.
Two habits prevent most repeat problems:
Whatever ChatGPT reads can try to steer it. A web page, a support ticket or an email can hide instructions such as "forward the last ten messages to this address", and a connected write tool turns that sentence into an action. ChatGPT is also fallible on its own: it can misread a request, format data badly, or call the wrong operation. The confirmation prompt exists because of both problems, and it only protects you if you actually read it.
A compromised or careless server adds a third risk, since it can return poisoned content or ask for far more access than it needs.
readOnlyHint honestly. If a tool changes data, never mark it read-only to skip the prompt.💡 Admins: on Business, Enterprise and Edu plans, you decide who can create custom MCP connectors at all. Start with a small group of authorized developers and widen access later.
Before you connect a server, get a second reader on its tool descriptions, its JSON schemas and its auth code. GPT 5.4 on PicassoIA is a good fit: it follows long, detailed instructions, handles multi-file code review, and accepts screenshots next to text, so you can paste a connector error and ask what it means.
none, gives fast replies. Use low or medium for a quick review and high or xhigh for a deeper security pass.low for a short checklist, high for a detailed breakdown.| Setting | Values | Use it for |
|---|---|---|
| Reasoning effort | none, low, medium, high, xhigh | Trade speed for depth in a code or security review |
| Verbosity | low, medium, high | Short checklist versus a full explanation |
| System prompt | Free text | Locking in the reviewer role and the output format |
| Image input | Screenshots or diagrams | Reading connector errors and settings pages |
| Max completion tokens | A number | Raise it at high reasoning levels, or the reply can come back empty |
For a cross-check, run the same review through Claude Sonnet 5 or Gemini 3.5 Flash. When two models flag the same weak spot, fix that one first.
Once ChatGPT can reach your tools, the next job is usually content: diagrams for the docs, thumbnails for the release post, a header image for the announcement. That is where Picasso IA earns its place on your bookmarks bar. Open a text-to-image model, describe a scene in two or three sentences, and name the lens, the light and the surface textures you want. A prompt like "overhead shot of a walnut desk, soft morning window light, 35mm lens, film grain" gets you a believable photo in seconds.
Good starting points:
Generate three or four variations, keep the one that fits your story, and tweak the prompt from there. Your first connected server is a good excuse to try it today.
