ChatGPT connectors and Apps: what they do and what they cannot

Filed under Guide

By Gerald · 5 August 2026

A blue LAN cable plugged into the back of a router, representing a live connection between two systems

Half the guides to this feature say "connectors." The other half say "Apps." Both are talking about the same thing, and the confusion is not your fault: OpenAI renamed connectors to Apps in mid-December 2025, folding both the interactive-UI apps and the plain data connectors under one umbrella term, while leaving every existing connection working exactly as before.

If you searched "chatgpt connectors" and landed here wondering whether you are looking at an outdated feature, you are not. You are looking at the current one, under a name that has not fully replaced the old one in people's search habits yet, or in a lot of the content still ranking for it.

A connector makes ChatGPT capable of reading your actual data. It does not make ChatGPT reliably choose to use that data every time it would help.

Connectors, Apps, and why the name changed

Before the rename, "connector" meant a data source, your email, your files, a work tool, that ChatGPT could search and reference. "Plugin" and later "app" mostly referred to things with an interactive interface inside the chat. OpenAI folded both concepts into one word, Apps, as of the December 2025 change, specifically because the line between "reads your data" and "does something with a visible interface" had already blurred. A connector that shows you a formatted table of your calendar is arguably both.

Then in July 2026 OpenAI migrated its app directory into what it now calls the Plugin directory, with Apps remaining the term for integrations that connect ChatGPT or Codex to external data and actions, and plugins referring more specifically to enabling particular workflows inside ChatGPT. If that sounds like two renames in eight months, that is because it was. Expect the terminology to keep drifting for a while yet. This article uses "connector" for the plain read-and-write data link and "App" for anything with its own UI, since that is still the clearer split in practice even if OpenAI's own labeling keeps moving.

What the feature actually is

A pair of hands typing on a black keyboard against a plain blue desk surface — Setting up a connector takes a few minutes. Knowing when the model will actually reach for it takes longer to learn.

Underneath the naming, a connector is a bridge between ChatGPT and an external system, built on the Model Context Protocol, the same open standard Anthropic introduced for Claude. When you connect one, ChatGPT gets a defined set of tools it can call: read your files, search your tasks, look something up in a work app you use. What it can do is whatever the connector's builder chose to expose, nothing more.

This is worth being precise about, because the marketing language ("connect ChatGPT to your world") makes it sound more open-ended than it is. A connector to a note-taking app that only exposes a search tool cannot also create notes for you, even if the underlying product supports that, because the connector itself has to explicitly offer that capability.

Built-in options versus a custom one

ChatGPT ships with first-party connectors to a handful of major services, available depending on your plan and workspace settings, alongside the ability to add a custom one to any MCP-compatible server. The built-in ones are the easiest path if the tool you want is already on that list. The custom path exists for everything else, which by mid-2026 is a fast-growing list of products shipping their own MCP server, Flow among them.

If you are on a Business, Enterprise, or Edu workspace, an admin generally has to enable custom connectors at the workspace level before individual members can add one, and admin controls let the organization restrict which ones are permitted. On a personal Plus or Pro account, you control this yourself from the connector settings.

Adding a custom connector step by step

Get the connector's server URL from the product you want to connect. For a Flow account, this lives on the Account page under "MCP Connector," and generating one gives you a URL with the access token already embedded in the path.
In ChatGPT, open Settings, then Connectors, and choose the option to add a custom connector, sometimes labeled under a developer or beta section depending on your plan.
Paste the server URL. If the product uses a separate bearer token rather than an embedded one, you will be prompted for that credential in a separate field.
Confirm the connection. ChatGPT will typically list the tools the connector exposes, read tasks, create a note, search files, whatever the server defines, so you can see up front what you are granting access to.
Start a new chat and reference the connected data directly ("what is overdue on my board") to confirm the model is actually calling the tool rather than guessing.

That last step matters more than it sounds like it should, which brings us to the part most guides skip.

When ChatGPT decides to use it, and when it ignores it

This is the actual gap between the setup instructions and real use. Connecting a tool does not force the model to use it. ChatGPT decides, per message, whether a question is worth a tool call, based on the wording of your prompt and its own judgment about whether the connector is relevant.

In practice this means vague or conversational phrasing ("what should I focus on today") sometimes gets answered from general reasoning rather than an actual read of your connected data, especially early in a conversation before the model has established that the connector is relevant to what you are asking. Specific phrasing that references the connected system directly ("check my Flow board and tell me what's overdue") reliably triggers the tool call, because there is no ambiguity about which source should answer the question.

If a response feels generic when it should have been personal, that is usually this, not a broken connector. Rephrase to name the tool or the data source directly and try again before assuming something is wrong with the setup.

The limits people run into

Write access is not guaranteed, and it is not the same for every connector. Some expose read-only tools, some allow creating or updating records, and a small number allow deletion. Whatever the connector's builder chose to expose is what you get; there is no way to grant ChatGPT broader access than the connector itself defines.

Custom connectors are not available on every plan or in every context. Availability depends on your plan tier and, for workspace accounts, on whether an admin has enabled the capability. Free-tier access is the most restricted, with the built-in connectors typically getting priority over custom ones.

Mobile support lags behind desktop. Built-in connectors generally work across ChatGPT's apps, but custom connector setup itself is easiest to manage from the web interface, and some newer capabilities, including full developer-mode MCP connections, have rolled out to desktop and web first.

Deep research and agentic modes handle connectors differently from a normal chat. A deep research task will proactively pull from connected sources as part of its multi-step process, while a standard chat message calls a connector only when the model judges it relevant to that specific turn. If you want guaranteed use of a connector, a research-style request is more reliable than a casual question.

The underlying protocol is MCP, not a ChatGPT-specific standard. This matters practically: a server built to the MCP spec can, in principle, work with ChatGPT, Claude, and other MCP-compatible clients without being rebuilt for each one, though how a client surfaces those tools and when it chooses to call them still varies by product.

Frequently asked questions

Do ChatGPT connectors work on the free plan? Access is more limited on the free tier than on Plus, Pro, or paid workspace plans, and custom connectors in particular are typically gated to paid tiers. Check your current plan's settings page for the specific list available to you, since this has changed more than once.

Can a ChatGPT connector write data, or only read it? It depends entirely on what the specific connector exposes. Some are read-only by design, others support creating or updating records, and a smaller number support deletion. There is no universal answer, check the tool list ChatGPT shows you when you add the connector.

Why does ChatGPT ignore my connector and answer from memory? The model decides per message whether a question warrants a tool call, and vague phrasing often does not trigger one. Reference the connected source directly in your prompt rather than asking a general question and expecting it to infer that a connector applies.

Is a ChatGPT connector the same thing as an MCP server? A ChatGPT connector is built on MCP, the Model Context Protocol. The connector is the specific configured connection inside ChatGPT; the MCP server is the piece of software on the other end that defines what tools are available. The same MCP server can, in principle, power connections in Claude or another MCP-compatible client as well.

Are connectors available in the mobile app? Built-in connectors are generally available across platforms, but managing and adding custom connectors is more fully supported on desktop and web, with some newer developer-mode capabilities rolling out there first.