ChatGPT Connectors in 2026: What They Are Now, How to Add One, and How to Use MCP

Go looking for Connectors in ChatGPT's settings today, and you won't

find them. OpenAI renamed the feature in December 2025 and moved it into

a new Plugin Directory in July 2026, so the old Settings \> Connectors

path is gone. The names are the easy part. The expensive part is the

plan question, because whether ChatGPT can change anything through a

custom MCP server depends on your ChatGPT plan, workspace permissions,

and the tools enabled for your account.

This guide covers what ChatGPT connectors are now, how to add one, how

to add an MCP server in developer mode, and what it takes to let ChatGPT

make changes in an ad account.

What are ChatGPT connectors?

**ChatGPT connectors let ChatGPT search, read, and in some cases act

inside tools you already use, such as Google Drive or Slack.** OpenAI

now calls them ChatGPT apps, and you install them from the Plugin

Directory under Settings \> Plugins. An app can only reach what your own

connected account can already open.

In practice, you ask "Find my project brief in Google Drive and

summarize the next steps" and get an answer drawn from the file itself.

You call an app with @ or the + menu, and ChatGPT may ask for approval

before reading or acting. Permission settings change when ChatGPT asks,

not what it can reach; in OpenAI's words, "App permissions do not

grant an app new access."

For ad teams, one distinction matters. A ChatGPT connector reads from

the service at the moment you ask, which makes it a different thing from

the ETL-style connectors that copy ad data on a schedule (some workspace

apps can also be set up by an admin to index content in advance).

Connectors, apps, and plugins: what the names mean in 2026

**Here is the sequence, dated from OpenAI's own [ChatGPT release

notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)

and help center.**

Date Change What it means for you

• ----------------------- ----------------------- -----------------------

March 23, 2023 OpenAI announces An earlier feature;

experimental today's ChatGPT

"plugins" plugins reuse the name

only.

May 8, 2025 GitHub connector Connectors appear as

arrives in deep deep research sources.

research

June 4, 2025 Custom connectors via The origin of custom

MCP, remote server MCP apps.

required, deep research

only

August 11, 2025 More connectors for The 2025 menu path,

Plus and Pro in chat, since replaced.

under Settings \>

Connectors

December 18, 2025 App directory launches; "Connector" becomes

connectors "now appear the legacy term.

in the directory as

apps"

July 9, 2026 Plugin Directory Settings \> Plugins is

replaces the App the current path.

Directory; plugins

bundle apps, skills,

and app templates

December 11, 2026 Custom GPTs retire and Integrations built on

(scheduled; Feb 11, migrate to plugins custom GPT actions need

2027 for approved a new home.

Enterprise deferrals)

Today there are two layers. An app connects ChatGPT to one service.

A plugin packages a workflow around it and can include apps, skills, or

both, plus app templates an administrator configures for a workspace,

and it works in Codex as well as ChatGPT. So "ChatGPT integrations"

can mean either the app or the plugin wrapped around it.

The last row matters for anyone piping ad data into custom GPTs: [OpenAI

says](https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq)

custom GPT actions "do not transfer through the migration workflow"

and that rebuilding the connection "may require a custom MCP server."

Which ChatGPT plans support connectors and custom MCP

Custom MCP access depends on your plan and workspace permissions.

For team workflows that need write actions, use a Business, Enterprise,

or Edu workspace with the necessary administrator permissions. On

individual plans, check the developer-mode options and actions available

in your account before building a workflow around them.

Plan Directory apps Custom MCP apps Write actions on

(developer mode) custom MCP

• ----------------- ----------------- ----------------- -----------------

Free and Go Select plugins; Not listed by No

no plugin OpenAI

extensions

Plus Yes, connected Check account Not confirmed

individually availability and

enabled actions

Pro Yes, connected Yes Read/fetch; check

individually enabled actions

before relying on

writes

Business Many on by Admins and owners Yes (beta)

default only

Enterprise and Admin-selected Admins, plus Yes (beta),

Edu set members granted per-action

via RBAC controls

Sources: [Developer mode and MCP apps in

ChatGPT](https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt),

[ChatGPT developer

mode](https://developers.openai.com/api/docs/guides/developer-mode),

[Plugins in

ChatGPT](https://chatgpt.com/features/plugins/).

**For a team workflow that needs custom MCP write actions, use a

Business, Enterprise, or Edu workspace and confirm the app's

permissions before running it.**

Data handling splits too: OpenAI does not train on connector data from

Business, Enterprise, and Edu by default, but may on Free, Plus, Go, and

Pro if "Improve the model for everyone" is on. For tools that spend

money, the plan only decides whether writes are possible; whether a

given write should happen is a job for the tool's own [approval gates

before budget

changes](https://syntermedia.ai/blog/how-synter-safety-gates-protect-ad-spend).

How to connect an app in ChatGPT

Install and connect from Plugins

OpenAI's current steps, from [Connected apps in

ChatGPT](https://help.openai.com/en/articles/11487775-connected-apps-in-chatgpt):

1. Open Settings \> Plugins and select the app, or the plugin that

includes it. Very old mobile versions may show Apps instead.

1. Select Install plugin if it appears.

1. Select Connect if prompted.

1. Sign in to the account that already has the information, such as the

Google account that owns your campaign briefs.

1. Review the requested permissions and finish the provider's

authorization.

Set app permissions

The [Managing app

permissions](https://help.openai.com/en/articles/20001495-managing-app-permissions-in-chatgpt)

article defines four options:

• Always ask. ChatGPT asks before reading or changing anything.

• Allow read actions. It reads freely but asks before changes.

• Allow low-risk actions. Low-risk actions run automatically;

higher-risk ones may need confirmation or be denied.

• Allow all actions. Supported actions run without prompts, which

OpenAI labels as elevated risk.

"Allow all actions" can appear for an individual app or connected

account, but not on the standard account-wide or workspace-wide

selectors. For anything connected to a budget, "Allow read actions" is

the sensible default: reporting questions flow, and every change still

stops for a human.

Workspace plans: what your admin controls

**In a Business, Enterprise, or Edu workspace, six separate settings

shape what you can do with an app:**

• Plugin installation settings (is the plugin available)

• Workspace app settings (can the underlying app be used)

• Enterprise and Edu role controls (which members get it)

• Provider authorization (which external account is connected)

• Action controls (which actions exist)

• App permissions (when ChatGPT asks first)

Business workspaces start with many apps enabled, while new Enterprise

and Edu workspaces start with a selected set, so an app showing

"Disabled by admin" is fixed by a workspace admin in Workspace

settings rather than in your browser.

How to add an MCP server to ChatGPT (developer mode)

**The Model Context Protocol is "an open-source standard for connecting

AI applications to external systems."** Developer mode is how OpenAI

documents adding your own MCP server, which is what most people mean by

ChatGPT MCP or ChatGPT custom connectors. MCP defines [two standard

transports](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports),

stdio (the client launches the server as a local subprocess) and

Streamable HTTP (messages are HTTP POSTs to one endpoint), and developer

mode connects over HTTP, supporting streaming HTTP and SSE.

Turn on developer mode

The toggle lives in one of two places, depending on your account and

rollout:

• **For individual accounts: check Settings \> Security and login \>

Developer mode or Settings \> Apps \> Advanced settings.**

• **For workspaces: admins can enable developer mode when creating an

app under Workspace settings \> Apps \> Create. Enterprise and Edu

members need administrator-granted access and can then enable it in

Settings \> Apps \> Advanced settings.**

OpenAI's developer guide warns about "prompt injections and other

risks, model mistakes on write actions that could destroy data, and

malicious MCPs that attempt to steal information." The [OWASP Top 10

for LLM

Applications](https://genai.owasp.org/llmrisk/llm01-prompt-injection/)

explains the mechanism: indirect injection happens when a model

"accepts input from external sources, such as websites or files."

Connect servers whose operator you can name and contact.

Create the app and point it at a remote MCP URL

Here is the flow with Synter's hosted server, an OAuth-protected

Streamable HTTP endpoint, as the worked example; Synter's own install

page follows the same sequence.

1. Go to Plugins and select + to create a developer-mode app.

Workspace admins can also use **Workspace settings \> Apps \>

Create**.

1. Name it Synter and enter the remote MCP server URL:

<https://mcp.synterai.com>.

1. Choose OAuth. Developer mode supports OAuth, no authentication,

or a mix of the two.

1. Select Scan Tools if shown, finish sign-in in the browser (never

paste an API key or token into the chat), and wait for tool

discovery.

1. Select Create. The app appears as a draft labeled Dev.

1. Open a new chat, select Synter, and run the install guide's

[first-read

prompt](https://syntermedia.ai/installation-instructions). Check

runtime pricing and approve any metered call first. Confirm the

workspace and advertiser account, then request a read-only report

with explicit dates and timezone. A saved app or tool list alone

does not prove the connection works; verify an authenticated tool

response. If access is missing or expired, fix that connection

before continuing.

Two details matter once it runs. Write actions require confirmation

by default, and ChatGPT treats any tool without a readOnlyHint

annotation as a write; a remembered approval lasts only for that

conversation. Business, Enterprise, and Edu admins can refresh an

existing app's tool list from workspace app settings without recreating

it; they need permission to manage the app and a valid connection.

Why ChatGPT cannot launch a local MCP server

A stdio server only works when the client launches it as a subprocess on

the same machine, and ChatGPT on the web does not do that. OpenAI's

help center is direct: "ChatGPT connects to remote MCP servers." For a

private or on-premises server, OpenAI's Secure MCP Tunnel runs a client

inside your network that opens an outbound HTTPS connection, with no

inbound ports.

Google's official Google Ads MCP server shows the gap. It runs locally

over stdio, or over HTTP if you deploy it to Cloud Run, so ChatGPT

can't reach it until it is hosted remotely or exposed through a tunnel.

The exception is the desktop app, where plugins can include local MCP

apps that do not carry over to web or mobile.

Best ChatGPT connectors for marketing and paid media teams

**As of October 3, 2026, the Plugin Directory has no Google-built Google

Ads app and no Meta-built Meta Ads app; searches return third-party

listings.** Both platforms publish official MCP servers, which you add

as custom apps. For a wider comparison, see our guide to [MCP servers

for each ad

network](https://syntermedia.ai/blog/best-ad-platform-mcp-servers).

Tool What ChatGPT can Read or write How it connects

do

• -------------------------------------------------------------------------------------------- ----------------- ----------------- ------------------------------

Google Drive Search Docs, Both Directory app

Sheets, and

Slides; create,

update, or share

files with the

right permissions

HubSpot Analyze CRM data, Both Directory app

such as which

campaigns led to

the highest-value

deals

Salesforce Listed as Both Directory app

"Manage

Salesforce

records"

Canva Turn an outline Creates designs Directory app

into a pitch deck

or other design

Windsor.ai Pull data from Both, each change Directory app

connected confirmed

sources; pause or

enable campaigns

and set budgets

on supported

platforms

Coupler.io Query ad, sales, Reads, per its Directory app

and finance data listing

from connected

data flows

ChatGPT Ads Manager (OpenAI) Create and manage Both, each change Directory plugin

ChatGPT ads; no confirmed

revenue or ROAS

reporting

Meta Ads (Meta's ads MCP) Reporting, Both Custom MCP:

campaign and ad https://mcp.facebook.com/ads

creation,

audiences; new

ads start paused

Google Ads (Google's MCP) GAQL reporting Read only Custom MCP, once hosted or

and account tunneled

discovery

Synter Reporting and Both; writes Custom MCP

supported changes depend on plan

across 27 ad (see above)

platforms

[Meta's server also enforces its own

rules](https://www.facebook.com/business/help/1456422242197840) on top

of ChatGPT's: anyone with full control of the business portfolio can

block actions per ad account or cap budgets, and Meta's example blocks

any budget above USD 1,000. Whatever you choose, pick one write path per

ad account, because two servers editing the same campaigns means two

rulebooks and two change histories.

Running paid media from ChatGPT with Synter

**We built Synter as an AI agent operator for ads, and ChatGPT connects

to it through our hosted MCP server as a custom developer-mode app. It

isn't in the Plugin Directory.** Behind one connection, our [Universal

Ads MCP](https://syntermedia.ai/mcp) reaches 27 ad

platforms, including Google Ads, Meta Ads, LinkedIn Ads, and OpenAI Ads,

though available actions vary by platform, account permissions, and

deployment scope.

On Business, an admin or owner creates and publishes the app for the

workspace.

If your ChatGPT plan cannot run write-capable custom apps, the same

endpoint works in Claude, Cursor, Codex, and Amp.

Take two of the example prompts on our MCP page: "Compare CPA and spend

velocity across Meta, TikTok, and Reddit over the last 30 days," then

pause LinkedIn creatives where frequency is above 4.5, and CTR has

dropped below 0.45%. The comparison returns data, while the pause stops

at ChatGPT's write confirmation by default and then meets Synter's own

controls:

• Approval modes. Review-required mode holds launches and edits for

a human, while auto-pilot runs inside guardrails, and the mode is

configurable per workspace and per campaign.

• Spend caps before the API call. Hard caps are enforced before

anything reaches the ad platform.

• Rollback. The change log can restore the last 10 changes per

entity, including budgets, bids, and targeting.

• Separate authorizations. Connecting ChatGPT to Synter and

connecting Synter to an ad account are two OAuth grants, and

installing the connection does not authorize spend.

There is a neat loop here for teams running [ChatGPT

Ads](https://syntermedia.ai/blog/chatgpt-ads-openai-ads-guide).

OpenAI's ChatGPT Ads Manager plugin manages those campaigns but

doesn't report revenue or ROAS. Our pull_openai_ads_performance tool

reports OpenAI Ads alongside the other platforms you connect, so you can

judge ChatGPT's own ad platform against Google and Meta from inside

ChatGPT.

Fixing "error fetching connectors" and other common problems

**[OpenAI's troubleshooting

article](https://help.openai.com/en/articles/20001497-troubleshooting-plugins-apps-in-chatgpt)

does not mention "error fetching connectors" by name.** Treat it as

ChatGPT failing to load your apps, and work through the checks OpenAI

does document, in order:

1. Look in the right place. Connections live under Settings \>

Plugins, or Apps on some accounts. Open the plugin details and

refresh before trusting a directory status.

1. Confirm the account. Check that you are in the intended account

or workspace, and that the app supports your plan and region.

1. "Disabled by admin." In a managed workspace, an administrator

reviews the app in Workspace settings; on a personal account, the

message alone does not mean anyone blocked it.

1. Re-authorize. If you see "All permissions are required,"

select Reconnect and approve every permission (Select all on

Google's screen).

1. Connected but not working. Start a new chat. If that fails, stop

reconnecting and check OpenAI's status page.

1. Reads work, writes fail. The app may not support the action, the

provider scope may be missing, an admin may have disabled it, or

ChatGPT may be waiting for your approval.

Custom MCP apps add three failure modes. If the OAuth provider does

not issue refresh tokens (offline_access), ChatGPT can lose access when

the first authorization expires. If the server's tools change after an

admin approves the app, ChatGPT keeps a frozen snapshot, and calls can

error until an admin refreshes, with no prompt telling anyone to. And

agent mode ignores custom apps, deep research uses them for read and

fetch only, and company knowledge only includes apps with search or

fetch tools.

Connectors are now apps, and the directory covers most everyday reading

plus a growing set of actions. For team workflows with custom MCP

writes, use a Business, Enterprise, or Edu workspace and check the

app's tools and permissions before relying on writes. To connect an ad

account, follow [Synter's ChatGPT install

steps](https://syntermedia.ai/installation-instructions)

and run the first-read prompt before approving a single change.

FAQ

What connectors should you use with ChatGPT?

Start with the apps that hold what you already ask about: Google Drive

or SharePoint for files, your CRM, and Slack or Teams. For ad accounts,

decide first whether you need reporting or changes, since that sets the

plan you need.

What apps can ChatGPT connect to?

Anything in the Plugin Directory, from Google Drive and Google Calendar

to CRMs and design tools, plus any remote MCP server you add in

developer mode on an eligible plan.

Which MCP server is best for ChatGPT?

The one whose operator you trust with the access it asks for. For ad

accounts, compare read versus write access, platform coverage, and the

controls enforced before a change lands.

Can I run MCP apps inside ChatGPT?

Yes. Developer-mode MCP apps run on ChatGPT web, not mobile, and ChatGPT

MCP servers you add yourself must be reachable remotely or through

Secure MCP Tunnel.

Do ChatGPT connectors cost anything?

Select plugins are available on every plan, though some require a paid

plan, and app use can count toward both your plan's limits and the

provider's own limits.

Can ChatGPT connectors take actions?

Yes, where the app supports them and your permissions allow it. For team

workflows, use a Business, Enterprise, or Edu workspace with the

required access. On individual plans, check which actions your account

exposes before relying on writes.