All documentation
On this page
Set-up documentation
Connect your mailbox to ChatGPT
ChatGPT Plus is enough: you switch developer mode on yourself, with no admin. We sent real email through it. Business and Enterprise need an admin first.
Which ChatGPT plan you need
Read this first, because it decides how much work the rest of the page is. The short version: ChatGPT Plus is enough, at the time of writing. You enable developer mode yourself and no administrator is involved. On a Business or Enterprise workspace an admin has to open the door before you can start.
We are stating the Plus case plainly because we tested it and because OpenAI's own documentation does not promise it. On we connected this server to a ChatGPT Plus account, the entry paid tier, in a web browser with developer mode switched on, and asked it to send an email. The message arrived, from the right address, with the sending account's own signature on it. Reading, searching and sending all worked.
That is more than OpenAI describes, on two counts. Their FAQ says full MCP is a Business and Enterprise/Edu feature, and that Pro accounts get read and fetch. It does not mention Plus at all:
Pro users can build apps using the Apps SDK. Full MCP is only available to Business and Enterprise/Edu users, currently. Pro users can connect MCPs with read/fetch permissions in developer mode.
OpenAI, Developer mode and MCP apps in ChatGPT
Both of those things are true as written, and we would rather show you the disagreement than pick the half that suits us. What we measured is what a Plus account did on the day we tried it. What OpenAI has written is what they currently commit to. If you are choosing a plan on the strength of this page, note that "currently" is doing real work in that sentence: start on Plus, because it is the cheapest way to find out, and do not commit a business process to a capability the vendor has not written down.
| Plan | Who switches developer mode on | What we have seen it do |
|---|---|---|
| Plus | You do, in your own settings, and it has to stay on. | Read, search and send. Tested 1 September 2026. |
| Pro | You do, the same way as on Plus. | Not tested by us. OpenAI documents read and fetch here, and Pro carries everything Plus has. |
| Business | An admin or owner, for their own account only. They cannot enable it for a member. | Read and send. This is the tier OpenAI documents for full MCP. |
| Enterprise and Edu | An admin, who can also grant it to named members through role-based access control. | Read and send, with per-action controls an admin sets. |
| Free | Not established. | We have not tested it and OpenAI does not name it. Assume no until somebody checks. |
Web only, and still in beta
OpenAI answers this one directly: MCP apps are not available on mobile. Set this up in a browser, and expect the mailbox to be missing from the phone app. They also describe full MCP support, including modify and write actions, as rolling out in beta, with functionality, interface and permissions liable to change. Take the screen names below as accurate on rather than as permanent.
Turn developer mode on
Nothing below exists until this is done, and the screens simply do not appear rather than telling you why. Where the toggle lives depends on your plan.
-
Plus and Pro
Your own settings, under Apps
You enable it for yourself and nobody has to approve it. OpenAI's note for personal tiers is worth acting on: you must keep developer mode enabled to use custom apps, so this is not a switch you flip once and forget. If the mailbox disappears from your tools list later, check here before you assume anything is broken.
-
Business
Workspace Settings, Permissions and Roles
Then Connected Data, and the Developer mode / Create custom MCP connectors toggle. Each admin or owner has to turn it on for themselves: switching it on does not apply it to the other admins in the workspace.
-
Enterprise and Edu
Settings, Apps, Advanced Settings
Admins and owners toggle it from their own user settings. They can also use role-based access control to grant developer mode to a chosen set of members, who then find the same toggle in their own settings once access has been given.
On a workspace plan you can also enable it for yourself at the moment you create an app, from Workspace settings, Apps, Create.
Create the app
-
Turn developer mode on
As above: yourself on Plus or Pro, an admin on a workspace plan. Nothing below exists until this is done, and the screens simply do not appear rather than telling you why.
-
Create the app
On a personal plan, and for an authorised member on a workspace plan, this is your own Settings, then Apps, then Create. Admins and owners can instead go to Workspace settings, then Apps, then Create. It asks for the endpoint and some metadata.
Endpoint The same for every accounthttps://mcp.mailbox-mcp.com/db/mcp/YOUR-MAILBOX-KEYWrite the description for the model rather than for a person. It is one of the things ChatGPT reads when deciding whether this app is relevant to a question, and a vague one is the usual reason an app is connected and never used.
-
Scan tools and authorise
Pick the authentication mechanism, then click Scan Tools. Because this server uses OAuth, an authorisation prompt appears: it sends you to the control panel to sign in and choose which mailbox this app should open. Wait for the scan to finish, then click Create.
There is no client ID and no secret to paste. The server registers its own client, which is covered in full on the Claude.ai page.
-
Test it, and on a workspace plan publish it
The app appears in your own settings under Enabled Apps with a Dev label beside its name. On a personal plan that is the finish line: it is yours, it works, and there is nothing to publish. On a workspace plan it also appears as a draft under Workspace Settings, Apps, Drafts, where an admin can publish it to everyone.
The refresh-token check
OpenAI asks you to check one thing about any OAuth server you attach, and
it is worth knowing the answer for this one before you go looking. Their
guidance is to confirm the provider issues refresh tokens, and to check its
discovery document for offline_access. Their warning is
specific:
If OAuth is configured without
OpenAI, Developer mode and MCP apps in ChatGPToffline_access, ChatGPT may lose access after the original authorization expires because refresh-token renewal may be unavailable, and users may need to reauthenticate.
This server passes both halves of that check, and you can read it yourself
rather than take our word for it. Checked against the live discovery
document on :
grant_types_supported lists authorization_code
and refresh_token, and scopes_supported lists
mcp and offline_access. So refresh tokens are
issued, and the literal scope name their check looks for is there.
This one used to fail, and the page said so
Until the discovery
document advertised mcp and nothing else. Refresh tokens
were always issued, so the connection would very likely have survived
anyway, but OpenAI's check reads the scope list and ours did not name
what it was looking for. It does now. The page carried the gap as a known
failure for the day it existed, which is the same standard everything
else here is held to.
Sending email from ChatGPT works, including on Plus
A fair amount of public commentary says MCP servers in ChatGPT are read-only, or that only search and fetch are supported. That was true of an earlier connector model and it is not true here. We have now tested it twice, on different tiers, because the second test contradicted the documentation.
-
30 August 2026
A workspace account, read and send
We connected a live mailbox through this route, read from one address and sent from another. Both worked, which is what OpenAI's documentation predicts for this tier.
-
1 September 2026
A Plus account, in a browser, sending real mail
One plain instruction of this shape, with no tool named and no folder named: send an email from [identity] to [address] with a really good morning message. ChatGPT chose the send tool, wrote the message, and it was delivered from the right address with that account's own HTML signature on it. OpenAI's FAQ does not mention Plus, and gives even Pro read and fetch only. A write tool ran on the entry paid tier and real mail arrived.
So the write tools are there: send, reply, forward, draft, rewrite a draft, move, flag and delete. They are the same tools every other client gets, because the server cannot tell the clients apart. ChatGPT asks for confirmation before some of them, based on the app's permissions and the context, and OpenAI says some especially risky actions are blocked rather than offered for approval.
Do not plan a business process around the Plus result
We are reporting a measurement, not a commitment. OpenAI does not name Plus in its MCP documentation at all, gives even Pro read and fetch, and describes the whole feature as a beta whose permissions may change. If sending from a personal plan matters to you, that is a reason to be somewhere the capability is actually written down, or to use a client without the ambiguity: Claude.ai sends on any plan and always has.
Testing, then publishing to a workspace
Test it before anybody else sees it: open a new chat, select the app from the tools menu or name it in your prompt, and ask for something only your mailbox can answer. On a personal plan this is the whole of the remaining job, and the publishing part below does not apply to you.
Try this first
List my five most recent emails with sender and subject, then tell me which ones are waiting on a reply from me.
Nothing is marked read by asking: reading through this server never sets the seen flag, so a test pass leaves your unread markers exactly as they were.
On a workspace plan, only admins and owners can publish. Workspace settings, Apps, Drafts, then Publish, reviewing the safety warnings, which are worth reading properly for a server that can send mail in somebody's name. Published apps appear in the workspace's approved connector list with a custom label.
On Business, publishing is one-way at launch
OpenAI states that on Business plans an app cannot be updated after publishing, and that changing its tools or metadata means recreating and republishing it. Enterprise and Edu have more control: an admin can configure which actions the app may take before and after publishing, and pull updated action definitions with a Refresh button, where new actions arrive disabled by default.
That last point is worth using rather than skipping. If this app exists to triage a shared inbox, deselect the sending actions and it cannot send, whatever anybody types at it.
The mailbox side is free to try: 5 MCP calls a day with no card, on any mailbox. On ChatGPT Plus you can have this working in about five minutes without asking anybody for permission.
Where to go next
If you are on a Business plan and are not an admin, or you want the mailbox on your phone, Claude.ai connects the same mailbox in four steps on any plan. The native connectors comparison covers what ChatGPT's own Gmail and Outlook apps do without any of this, which is a fair question to answer before asking an administrator for anything.
- Whether a personal plan keeps the ability to call write tools. We sent real email from a Plus account on 1 September 2026, but OpenAI's MCP documentation does not mention Plus at all and gives even Pro read and fetch only, so our result is well beyond what they promise and could be withdrawn without notice.
- Whether ChatGPT Free reaches a custom MCP server at all. We have not tested it and OpenAI does not name it. Pro is untested by us too, though it carries everything Plus has.
- The exact screen path on a personal plan, both to the developer mode toggle and to the Create screen. OpenAI documents the Business and Enterprise paths and not this one, and our own control panel describes a different route again, so we say where it lives rather than naming screens we cannot source.
Read 1 September 2026 Every value on this page came off OpenAI: developer mode and MCP apps in ChatGPT on that day. Full MCP support is in beta, so the screen names may move. If it has changed since, tell us and we will read it again.