A founder types "post this launch thread" into ChatGPT. If the connected tool can publish, it may do exactly that, on whichever account is linked, before anyone reads the text. That is the case for a draft-first test. Ask for a draft, read the preview, confirm nothing went out, and only then approve anything.
ChatGPT connectors let ChatGPT call tools on another service. A custom connector is a URL that points ChatGPT at a tool server, and you decide whether that server can only read or can also act. A connector that can act deserves a dry run before it gets a real job.
We use one worked example, Saymany, an MCP connector that posts to social networks from Claude or ChatGPT. The steps work for any server you might add. Where a bigger scheduling tool fits better, we say so.
Table of Contents
- What ChatGPT Connectors Actually Are
- How Custom Connectors Use MCP
- Which ChatGPT Plans Can Add Custom Connectors?
- Add a Connector Step by Step
- Check What a Connector Can Do Before You Trust It
- Are ChatGPT Connectors the Same as Apps or Plugins?
- Frequently Asked Questions
What ChatGPT Connectors Actually Are
A connector lets ChatGPT read from or act in an outside service. That service might be a calendar, a file store or a social posting tool. Without a connector, ChatGPT only knows what you type and what it learned in training. With one, it can fetch live information or do something on your behalf.
The label varies. OpenAI's help pages have described this area under several names, and the word in your account may differ from the word in a blog post. OpenAI's help article on connectors in ChatGPT is the place to check current wording.
Two kinds you will meet
- Built-in or partner connectors. OpenAI or a partner built them. You pick them from a list, and the sign-in is usually handled for you.
- Custom connectors. You point ChatGPT at a server address yourself. The server decides which tools you get.
Read-only and write
Connectors also differ by what they can do. A read-only connector fetches information, such as a document or a report. A write connector takes actions, such as sending an email or publishing a post.
The risk is different. A bad read can give you a wrong answer. A bad write can put something public or change real data. Keep that difference in mind, because the safety section below depends on it.
How Custom Connectors Use MCP
Custom connectors use MCP, short for Model Context Protocol. It is an open standard that lets an AI app discover and call tools on a remote server. The AI app is the client. The server describes what it can do, and the client uses that description.
OpenAI documents custom MCP servers in its help article on Developer Mode, apps and full MCP connectors. The article labels the feature as beta, so expect changes.
What happens when you ask for something
- ChatGPT reads the list of tools the server exposes, with a short description of each.
- You write a request in plain language.
- ChatGPT picks a tool that fits and sends the request with the details it needs.
- The server does the work and returns a result.
- ChatGPT shows you that result in the conversation.
Some setups ask you to confirm before a tool runs. That confirmation is a useful brake, so do not rush past it.
Why a connector is just a URL
The server address is the whole setup. You paste it in, sign in if asked, and the tools appear. For ChatGPT, the address generally needs to be a public HTTPS endpoint, which is why a server on your laptop will not work.
Claude supports the same kind of connector. One MCP server can therefore work in both apps. Anthropic's help pages list Claude's current menu names.
Which ChatGPT Plans Can Add Custom Connectors?
The plans that can add custom MCP connectors are the ones listed in OpenAI's Developer Mode article. That list is the one to trust, because the feature is in beta and eligibility can change. Your own settings give the fastest answer. If Developer Mode and a way to add a custom connector appear, your account has it. The connectors help article and the wider ChatGPT connectors article cover the rest.
Access depends on plan and workspace
On a team or business workspace, an admin may need to switch the feature on before anyone sees it. If the option is missing, ask your admin before you assume your plan is wrong. A connector can also connect and still expose only read tools, or none, when a workspace limits write actions.
Add a Connector Step by Step
Menus may change, and OpenAI labels this feature beta. The order of steps below is stable, and the names follow OpenAI's own article titles. The Developer Mode article has the current path through settings.
- Open Settings in ChatGPT and find the connectors area.
- Turn on Developer Mode, if your account asks for it.
- Choose the option to create a custom connector.
- Paste the MCP server URL.
- Approve the sign-in if the server asks for one.
- Open the tool list and read it before you use anything.
A worked example with Saymany
For Saymany, the connector URL is https://saymany.com/mcp. Paste it at step 4 and follow the sign-in that appears. The tool list should show tools for drafting, previewing, scheduling and publishing posts. Our ChatGPT setup guide walks through the exact screens, and the Claude setup guide covers the same URL there.
Run a draft-only test first
Do not ask for a live post on day one. Connect one low-risk account, such as a test channel or a spare Bluesky profile. Then paste a prompt like this:
Draft a post announcing our new pricing page, with this link:
https://example.com/pricing
Draft only. Do not publish or schedule anything.
Show me the preview for Bluesky before anything is sent.
A good preview shows the exact text for each network, the link as it will appear, any media, and a clear sign that the post is still a draft. Here is a sketch of a pass and a fail. It is not a captured log, and the wording will differ by connector.
PASS
Draft for Bluesky (not sent)
"Our pricing page is live. See the plans: https://example.com/pricing"
Status: draft. Nothing has been published or scheduled.
FAIL
Published to Bluesky.
If the reply reads like the second block, stop and disconnect. A preview with no draft label, or a reply that says a post was published or scheduled, is a fail even if the text looks fine.
Now check the place that matters. Open the test account's profile feed in your browser, or the test channel in Telegram or Discord. Nothing new should be there. Then give the next instruction:
Looks good. Schedule it for ten minutes from now.
If your team uses approvals, require one and approve it. Ten minutes later, reload the same feed and confirm the post appeared once, with the text you approved.
Pro Tip: Put "draft only, do not publish" at the top of your first prompt. A clear instruction costs nothing, and you can drop it once you trust the flow.
To try this loop on Saymany, the 14-day trial needs no card, and the plans and pricing are public. Run the dry run there first, and give it a real job only if it passes.
Turn it off when you are done
To disconnect, open the same settings and remove the connector. If you signed in with an API key, also delete that key in the service itself. Removing a connector in ChatGPT stops the chat from calling it, but a key you created stays valid until you revoke it.
Check What a Connector Can Do Before You Trust It
A connector can only be as safe as the service behind it. Here is the failure to design against. You ask for a draft of a launch post. The connector exposes a publish tool, and the model calls it.
The post goes to the linked account, which is your main brand profile and not the test one, and the link has a typo. Each item below stops part of that story.
- Connect only services you trust and can name an owner for. Someone must answer for what the service does with your accounts.
- Read the full tool list and mark each tool as read-only or able to change things. A publish tool you spotted in advance is one you can guard against.
- Link a low-risk account first. A wrong-account post then lands on a test feed.
- Require a preview and an approval. The broken link gets caught before it goes out.
Sign-in and least access
Sign-in usually works in one of two ways. An approval screen, often called OAuth, lets you grant access in a browser. Tokens from it can expire, and you can revoke them. An API key is a secret string you paste in. It is simple, but it works like a password, so store it carefully and revoke it if it leaks.
Grant the least access that does the job. If you only need drafts, do not hand over publishing rights. Fewer permissions mean less damage from a wrong prompt.
A no-surprises habit
Use the same loop for any write connector. Draft first. Check the preview. Get a human to approve. Then publish. Know how to disconnect fast before you need to.
Why one prompt cannot fit every network
Each network has its own character limits, media rules, link handling and API permissions. An official API can reject a post even when your prompt was fine. So one brief often needs a separate version for each network, and you should see every version before it goes out.
We show a preview before anything goes out, and teams can require approval. Today we post to X (through bundle.social), Bluesky, Telegram and Discord. Instagram, Facebook, Threads, LinkedIn, TikTok, YouTube, Pinterest and Google Business Profile are rolling out as each platform's review completes. X posts are billed at X's own cost.
When another tool fits better
A bigger scheduling tool can be the better choice, and each one wins on something a chat cannot give you.
- Buffer and Later. Both are known for a visual calendar. If you plan a month of posts by dragging them around a grid, a chat gives you a list instead.
- Hootsuite. It is aimed at larger teams that want governance and reporting in one place. A chat connector will not replace that.
- Metricool. It is built around analytics and reporting. You post from a chat, but you would still study results somewhere else.
- Ayrshare. It is a social publishing API. If you are building posting into your own app, you want an API you call from code, not a chat connector.
A conversational connector fits best when you already live in the chat and want to skip a dashboard.
Are ChatGPT Connectors the Same as Apps or Plugins?
They are related, but the names have shifted. The original plugin system was an earlier idea. Today's connected apps and custom MCP connectors cover similar ground in a different way. Do not assume that an older plugin article describes what you will see now. OpenAI's current documentation uses different terms for different surfaces.
Two people can also see different labels for what looks like the same feature. Interface versions differ, and so do plans and workspace settings. One person may see "apps" where another sees "connectors." A third may need Developer Mode before a custom option appears at all.
A social posting connector sits in one place in this picture. It is a custom MCP connector you add by URL. It is not a built-in ChatGPT feature, and OpenAI does not ship it for you. That is why the setup steps above matter more than the label.
How to tell read-only from write-capable
Open the tool list and read the names and descriptions. Tools that fetch, search or list are read-only. Tools that create, send, schedule, publish or delete can change things. Then follow the safety checklist, with draft-only first and approval before anything goes out.
The draft-first habit exists to prevent one failure. A launch post goes live on the wrong account, with a broken link, before you ever saw it. A dry run and a preview catch that. Trust is earned by a connector that passes a dry run, not granted by a URL that connects.
Frequently Asked Questions
What are ChatGPT connectors?
They let ChatGPT read from or act in an outside service, such as a calendar, a file store or a social posting tool. OpenAI has used several names for this idea, including connectors, apps and plugins. Check OpenAI's help pages for the wording in your account.
How do I add a custom connector to ChatGPT?
Open your connector or app settings, choose the option to add a custom connector, and paste the MCP server URL. Approve the sign-in if asked, then read the tool list. Follow OpenAI's current help article for the exact path.
Is a ChatGPT MCP connector safe to use?
It is as safe as the service behind it and the access you grant. Connect only services you trust, read the tool list, and give the least access that works. For write tools, draft first, check the preview and approve before publishing.
Does Claude support the same connectors?
Yes. Claude supports the same kind of MCP connector, so one server can work in both apps. Anthropic's help pages list Claude's current menu names.
Why can I connect a server but not post?
Write actions are often limited more than read-only access. Check your plan, your workspace permissions and whether an admin must enable custom connectors. Also confirm the connector offers write tools at all.



