One User's MCP Hack Shows Why OpenAI Killed Unlimited Chats

What the developer actually built

The setup was not a jailbreak. It was OpenAI's own tooling, used in a way the company probably did not plan for.

According to his Hacker News post, the developer ran a unified MCP server on a Windows laptop and a VPS. Through it, ChatGPT could read and write files, run terminal commands, execute tests and manage processes. A computer-use agent added screenshots, mouse clicks and visual checks of browser or desktop apps.

The bridge was OpenAI's Secure MCP Tunnel, authenticated with an API key. That meant ordinary ChatGPT conversations could inspect repositories, edit code and verify changes, without separate paid API inference calls.

The workflow was the clever part. A stronger reasoning model wrote a plan and broke it into GitHub issues with acceptance criteria. Several separate chats, each on a cheaper model, picked up different issues and worked in parallel. GitHub acted as the task queue.

He even built a small dashboard to start and stop the MCP connection, watch CPU and memory, and follow a live feed of tool calls. It is a hobbyist project, but a disciplined one.

Why OpenAI is merging chat and Codex limits

The economics are hard to argue with once a chat window can run commands on your machine.

A chat message and an agentic coding session look similar in the interface but cost very different amounts to serve. An agent that reads files, runs tests and retries after errors can burn many times the compute of a question-and-answer exchange.

If one pool of usage is unlimited and the other is metered, heavy users will drift to the unlimited pool. That is not abuse in any malicious sense. It is exactly what the developer describes: making the best use of what OpenAI made available.

The company has not published a detailed rationale for the change, and the developer is clear that he does not seriously think his setup caused it. But the pattern is familiar. Flat-rate plans tend to survive only until a small share of users finds a way to turn them into infrastructure.

The bigger story: subscriptions are becoming usage contracts

For most of ChatGPT's life, the pitch was simple: pay a monthly fee, chat as much as you like. That worked while the product was mostly text in, text out.

Agents change the math. When a model can act on your computer, a single conversation can consume hours of compute. Every major AI subscription is now somewhere on the road from flat rate to metered usage, and OpenAI's move is the most visible step yet.

The developer's reaction is telling. He says unlimited chats were the only thing keeping him on an OpenAI subscription and that he prefers Claude models, especially Opus, for his work. Whether or not he leaves, the episode shows how quickly a pricing change can turn a loyal subscriber into a comparison shopper.

What it means for you

If you are on a ChatGPT plan in the US, UK or EU, expect your chat usage and any Codex usage to be counted together once the change lands. OpenAI has not published the new quotas, so anyone running agent-style workflows should assume their effective headroom will shrink.

The practical advice is to separate experimentation from production. Use metered API access for anything that runs unattended, and keep subscription chats for interactive work. That costs more on paper, but it makes the bill predictable.

The unlimited chat era is closing because agents made it too expensive to keep open. The developer's MCP setup was clever, well documented and entirely within the rules, which is exactly why it was never going to last. The next fight will be over how generously the merged limits are drawn.