Large Language Models

ChatGPT MCP Plus: Connector Limits, Not Working Fixes and Free Alternatives

ChatGPT Plus can talk to MCP servers, but the rules around plans, read and write access, rate limits and errors are confusing. This article sorts out what Plus allows, why connectors fail, and which free alternatives do similar work.

ChatGPT MCP Plus: Connector Limits, Not Working Fixes and Free Alternatives
Cristian Da Conceicao

You paste your MCP server address into ChatGPT, press Create, and get a red error banner. Or the connector saves fine, then every tool sits greyed out. Or it worked yesterday and today ChatGPT acts like it has never heard of your server. If that sounds familiar, you are in good company. ChatGPT MCP Plus: connector limits, not working fixes and free alternatives is exactly the pile of questions people keep asking, because the rules changed several times in a year and most write-ups contradict each other.

This article sorts out what a Plus subscription actually gives you for custom connectors, why they break, what to check in what order, and which free alternatives can do similar jobs when the connector route is a dead end. Where sources disagree, I say so instead of guessing. Everything here reflects what was published as of November 2026, and the ChatGPT interface moves fast, so treat menu names as approximate.

💡 Quick answer: Plus can use custom MCP connectors, but access, read or write permissions and rate limits depend on your account, your region of the rollout and the current beta rules. When one fails, the cause is usually the server, the sign-in token or a stale tool list, not your subscription.

What MCP Does in ChatGPT

The short version

MCP stands for Model Context Protocol. It is a standard way for an AI chat to call tools that live somewhere else: a calendar, a database, a project tracker, a notes app, your own scripts. Instead of copying data into the chat window, you point ChatGPT at a server, the server lists the tools it offers, and the model decides when to call them.

In ChatGPT this shows up as a connector (sometimes labelled an app or plugin, depending on the week and the rollout). You give it a public HTTPS address, pick an authentication mode, and ChatGPT reads the tool list.

Woman typing on a laptop at a bright kitchen table

Where connectors fit in daily work

Most people use connectors for three jobs:

  • Reading data from a service, such as pulling a document or a ticket list into the conversation.
  • Writing to a service, such as creating a task or posting a note.
  • Running a custom tool you built yourself, such as a script that checks a server or formats a report.

The first is the easiest to get working. The second is where permissions bite. The third is where most "not working" threads come from, because a home-built server rarely matches the spec on the first try.

ChatGPT Plus Connector Limits

Who gets custom connectors

Several third-party guides agree on the broad picture: custom MCP connectors are available on the paid tiers, including Plus, Pro, Business, Enterprise and Edu, and not on Free or Go. On Business and Enterprise plans, a workspace admin usually has to switch custom connectors on before anyone else sees the option.

For years the route went through Developer Mode, which was explicitly labelled a beta. A recent guide from November 2026 says you no longer need Developer Mode and describes adding servers from a plus button in a Plugins area instead. Other pages still describe the older path. If your settings do not match a tutorial, that mismatch is probably the reason, not something you did wrong.

Read versus write access on Plus

This is the most disputed point. Some vendor help pages say Plus and Pro users are limited to read or fetch only connectors. Another source says Plus gets full client support for all tools, read and write. A third says write and action features on Plus are restricted rather than absent.

I could not find one authoritative answer that every source agreed with, so test it yourself: add a harmless write tool (for example, "add a line to a scratch file") and see whether ChatGPT asks to run it or refuses outright.

Tidy rack of network patch panels with colored cables

Rate limits, agent mode and platform gaps

Reports describe a few other limits worth knowing before you build a workflow around connectors:

  • ChatGPT applies its own per-tool rate limits on top of whatever your server enforces.
  • One source says agent mode does not use custom MCP apps at all.
  • The same source says MCP apps work on the web only, not in every mobile client.
  • Workspaces can block connectors entirely, so a work account may behave differently from a personal one.
LimitWhat people reportWhat to do
Plan accessPaid plans only, Free and Go excludedCheck your plan page first
Write access on PlusSources disagreeTest with a harmless write tool
Rate limitsPer-tool caps on top of server capsBatch calls, return less data
Agent modeReported not to use custom appsUse a normal chat for connector work
MobileReported web onlyTest on desktop first
Workspace accountsAdmin must enable connectorsAsk the admin, or use a personal account

Why Connectors Stop Working

Error 424 or a TaskGroup message on creation

A common thread on the community forums: you try to create the connector and get a 424 or a "TaskGroup" error. Some posters say ChatGPT never even sends a request to their server. Restarting the MCP server and refreshing the connector page fixed it for several people. If your server logs show nothing arriving, the problem sits before your code: a wrong address, a blocked port, a tunnel that went to sleep, or a certificate ChatGPT does not trust.

Tools greyed out or missing

You add the connector and the tool list is empty, or every tool is greyed. Two causes show up again and again:

  1. A frozen snapshot. ChatGPT keeps a snapshot of the tools your server listed when you created the connector. If you renamed a tool or changed its inputs afterward, calls can fail until you refresh the connector or remove and re-add it.
  2. A spec problem. If the server returns the tool list in a shape ChatGPT does not accept, it shows nothing, with no helpful message.

Frustrated man in a gray hoodie looking at a laptop at night

Expired sign-in

For connectors that use OAuth, an expired or revoked token is one of the most common reasons a once-working connector goes quiet. The usual fix is to revoke access in the connected app, then authorize again inside ChatGPT. If the connector only fails after being idle for a day or so, suspect token expiry before anything else.

The "unsafe connector" warning

Some users report a warning that a connector is unsafe when it is connected without any authentication. That makes sense: an open server that anyone can call is a risk, and ChatGPT is cautious about it. Adding proper sign-in, even a simple one, usually clears the warning and is better practice anyway.

Plain old outages

OpenAI has posted incidents where Plus and Pro users saw elevated errors with connector tools. Before you spend an hour debugging your server, check the status page. If there is an incident open, wait it out.

Fix It in Ten Minutes

A checklist that works in order

Run these steps in order and stop when the connector works. The order matters: it goes from cheapest to most disruptive.

Wooden desk with a handwritten checklist and a laptop

  1. Check the status page. If connectors are degraded, stop here.
  2. Confirm your plan and role. Free, Go and some workspace roles cannot add custom connectors.
  3. Open the address in a browser or a test client. A healthy server should respond. A timeout means the problem is on your side.
  4. Use HTTPS with a valid certificate. Self-signed certificates and plain HTTP are frequent blockers.
  5. Refresh the connector. In the connector settings, look for a refresh option so ChatGPT pulls the current tool list.
  6. Remove and re-add it. Annoying, but it clears stale snapshots and broken sign-in state.
  7. Re-authorize. Revoke access on the other service, then connect again.
  8. Start a new chat. Old conversations can hold on to the previous tool list.
  9. Try another browser or a private window. Extensions sometimes break the OAuth popup.
  10. Test with a minimal server. If a tiny server with one tool works, the problem is in your tool definitions.

💡 Tip: Keep tool names short and stable, and write a one-sentence description for every tool. Models pick tools by their descriptions, so a vague description gets ignored even when the connector is healthy.

Test the server before blaming ChatGPT

A separate MCP inspector tool lets you connect to your server and call each tool by hand. If a tool fails there, no amount of settings tweaking in ChatGPT will help. If it works there but not in ChatGPT, compare the tool list, the sign-in mode and the address you pasted.

Also watch for the classic tunnel trap: a temporary public address that changes every time you restart your tunnel. ChatGPT still points at the old address, so the connector fails silently. Reserve a fixed address if you can.

Developer reconnecting a cable to a laptop in a bright workspace

Safety Before You Connect Anything

Treat every connector like a house key

A connector gives a chat model the ability to read your data and, sometimes, to change it. A few habits keep that safe:

  • Start read only. Add write tools only after the read tools behave.
  • Grant the narrowest permission. If a tool only needs one folder, do not give it the whole drive.
  • Review what the model is about to send. ChatGPT typically shows tool calls for approval. Read them, especially for writes and deletes.
  • Never paste long-lived secrets into a chat to make a tool work. Use the sign-in flow the server supports.
  • Be careful with servers you did not write. A third-party server can return text that tries to steer the model. If a result tells the model to do something unrelated to your request, stop and check.

Older man with gray beard reading a printed page beside a laptop

Workspaces add another layer

If you use a work account, an admin may restrict which connectors can run, and who can create them. A connector that works on your personal Plus account may be blocked at work for good reason. Ask before you route company data through a server you built over a weekend.

Free Alternatives That Work Today

When the connector route is not worth it

Sometimes the honest answer is that you do not need a live connector. If your goal is "have an AI read this data and do something with it", you can export the data yourself and paste it into any capable chat model. You lose automation, but you gain reliability, and you skip the plan limits entirely.

Three colleagues gathered around a laptop at a round table

Models worth trying on PicassoIA

ChatGPT itself is not a model page on PicassoIA, so there is no ChatGPT connector to set up here. For similar writing, coding and reasoning work, these language models are on PicassoIA today, and several are labelled free on their pages:

ModelGood forLink
GPT 5.6 SolCoding tasks and long reasoningGPT 5.6 Sol
GPT 5.4Everyday writing, code and reasoningGPT 5.4
GPT 5 StructuredReplies in clean JSONGPT 5 Structured
Claude Sonnet 5Coding and automation workClaude Sonnet 5
Gemini 3.5 FlashFast chat and codeGemini 3.5 Flash
Kimi K2.6Agent style tasks and codeKimi K2.6
Deepseek v3.1Writing, code and reasoningDeepseek v3.1
Granite 4.1 8BLight chat and code, listed as freeGranite 4.1 8B

The older GPT-4o page is also listed as free, which makes it a low-risk place to test a prompt before you move it elsewhere.

What you give up, honestly

Be realistic about the trade. A pasted export is a snapshot, so the model cannot check live data or write back to your tools. If you need a model that acts on your systems all day, a connector, or a proper agent setup, remains the right tool. If you need drafts, summaries, code reviews or structured output from data you already have, a plain chat does the job with fewer moving parts.

How to Get Similar Results on PicassoIA

The connector workflow boils down to "give the model context, ask for a result". Here is how to reproduce the useful part with GPT 5.6 Sol.

Woman in a denim shirt working at a standing desk with two monitors

Step by step

  1. Open the model page for GPT 5.6 Sol.
  2. Export your data first. Copy the ticket list, log lines or document text that a connector would have fetched.
  3. State the job in one sentence at the top, for example: "Summarize these tickets by customer impact."
  4. Put the data in a clearly marked block, such as text between two lines of equals signs, so the model does not confuse it with instructions.
  5. Say what the output should look like: a table, a bullet list, a short email.
  6. Send, then refine. Ask for changes in the same chat so the context stays attached.

Parameter and prompt tips

  • Keep the context tight. Remove columns and lines you do not need. Short input gives sharper answers.
  • Ask for structure when another tool will read the result. GPT 5 Structured returns clean JSON, which is handy if you plan to feed the output into a script.
  • Split big jobs. Run one chat per data source instead of mixing five exports.
  • Compare two models on the same prompt. Try Claude Sonnet 5 and Gemini 3.5 Flash and keep the answer that reads best.

💡 Tip: If your connector kept failing, save the working prompt from your pasted version. Once the connector behaves, you will already know which instructions the model responds to.

Try It Yourself

Connectors are great when they work, and frustrating when they do not. You do not have to lose a day to a 424 error. Export the data, pick a model, and get your answer in a few minutes, then decide whether the connector is worth fixing later.

Open the model pages above on PicassoIA and run the same prompt through two or three of them. You will quickly see which one fits your writing, your code or your data. When you want a different kind of output, browse the full catalog at picassoia.com/en/all-models and try an image or video model on the same idea.

Share this article

Choose your language

FreeCreate AI Images and VideosFree, No Credit CardFree Generations to StartSeedance 2.5 Lite IncludedStart Free Now