I Am the Reason OpenAI Is Killing Unlimited Chats

Okay, probably not. But after what I've been doing with ChatGPT, I can understand why they might want to.

OpenAI recently announced changes that will kill unlimited ChatGPT chats and merge it with Codex usage. The unlimited chats I'd been using for development won't work quite the same way.

For the past few months, I've essentially been doing all my software development using unlmited ChatGPT conversations without burning through my Codex limits.

It started with MCP

ChatGPT's support for custom MCP servers and OpenAI's Secure MCP Tunnel meant I could expose tools running on my own machines directly to regular ChatGPT conversations.

So I started experimenting.

I built a unified MCP server that runs on my Windows laptop, with a similar setup on my VPS, and integrated a few open-source projects:

• Desktop Commander: Gives ChatGPT the ability to access files and folders, read and write code, execute terminal commands, run tests, and manage processes.
• CUA (Computer Use Agent): Adds computer-use capabilities, including screenshots, mouse interactions, clicking buttons, inspecting visual elements, and testing browser or desktop applications.

Suddenly ordinary ChatGPT converssations could do most of the practical development work I previously used Codex for.

They could inspect repositories, modify code, run commands, debug errors, and visually verify their changes.

Connecting everything through OpenAI's tunnel

The next challenge was making my local MCP servers accessible to ChatGPT.

OpenAI provides a Secure MCP Tunnel, which turned out to be exactly what I needed.

I connected my unified MCP gateway through the tunnel, authenticated the connection using an API key, and exposed my MCP tools directly to ChatGPT.

No custom cloud infrastructure was needed just to bridge ChatGPT to my laptop.

The tunnel and MCP integrations didn't require separate paid API inference calls. I used open-source tools, OpenAI's provided tunnel tooling, my existing ChatGPT subscription, and computers I already had.

I improved the integration

Of course, once I got everything working, I couldn't resist tinkering with the tools themselves.

I made several enhancements to the CUA integration to improve how ChatGPT interacted with my computer and how results appeared inside conversations.

One example was getting actual screenshots to render directly inside ChatGPT.

Instead of receiving a screenshot tool result that ChatGPT could inspect but I couldn't easily see, I added support for displaying captured screenshots through a custom MCP App viewer.

I also improved how tool-call results were handled, making them easier for me to read within ChatGPT chats

My fun little AI development workflow

Since ChatGPT chats cannot spawn or control subchats/subganets so I had to be a bit creative in order to do multi-agent development and orchestration.

I'd start with a more capable reasoning model, such as Astra in Codex or the limited Pro model in ChatGPT, and use it to create a detailed implementation plan.

Next, I'd ask it to break the plan into smaller, well-defined GitHub issues. Each issue contained enough context, requirements, and acceptance criteria that another AI agent could pick it up independently.

Essentially, I was turning a large development task into a queue of smaller jobs.

Then I'd open several separate ChatGPT conversations, each running Sol 6 High, and assign them different issues.

Because all those chats had access to my MCP server, they could actually work on the code.

They could read the project, edit files, run tests, inspect errors, and iterate until their assigned task was complete.

And at the time, I could use Sol High chats without hitting the kind of usage limits I encountered in Codex.

Sure, Sol High was slower than the premium models, but when usage was effectively unlimited, I was willing to put up with it.

I'd have multiple conversations working on separate tasks, while saving my more limited premium model usage for architecture, difficult decisions, and reviews.

It wasn't perfect. Parallel agents sharing a filesystem need coordination, and occasionally an issue would require a stronger model to intervene.

But for well-defined coding tasks, it worked surprisingly well.

I was essentially using ChatGPT as a pool of coding agents, with GitHub issues acting as the task queue.

I even built a dashboard for it

I also built a super lightweight interface where I can start or stop the entire MCP connection, monitor server health, track CPU and memory usage, and inspect a live feed of tool calls.

I can see when ChatGPT reads a file, executes a command, or invokes a computer-use tool.

The dashboard also tracks failures and lets me inspect recent activity.

OpenAI has now killed unlimited chats

As like all good things, this must also come to an end. I am sad that this is now going away.

Unlimited chats was the only thing holding me back to an OpenAI sub. I personally prefer the Claude models greatly and I am going to jump ship. Claude Opus in particular seems to be much smarter for my use cases and I also get more mileage out of a Claude sub vs a Codex sub.

To be clear, I am not breaking any rules.

I am simply making the best use of what OpenAI has made available to me: ChatGPT conversations, MCP support, secure tunnels, and the ability to connect my own tools.

Of course, I don't seriously think my setup had anything to do with OpenAI's decision. And although I am sad about not having unlimited chats anymore, but I am still glad about the extra usage i was able to squeeze out of my sub.

So, OpenAI, if you're reading this: I'm not sorry about that. Claude here i come.