I Am the Reason OpenAI Is Killing Unlimited Chats

Who Should Read

A developer who wants to run parallel coding agents on a ChatGPT subscription while saving Codex limits

Core Mechanics

• Local tools wired into regular ChatGPT chats
  • A custom MCP server plus OpenAI's Secure MCP Tunnel exposed tools on the author's own machine to ordinary conversations.
  • No custom cloud infrastructure was needed.
• Desktop Commander plus CUA
  • Desktop Commander handled file access, reading and writing code, terminal commands, tests, and process management.
  • CUA (Computer Use Agent) handled screenshots, mouse actions, and testing browser or desktop apps.
• No extra API cost
  • The setup used open-source tools, OpenAI's tunnel, the existing ChatGPT subscription, and computers already owned.
  • The tunnel and MCP integrations needed no separate paid API inference calls.
• One planner, many workers
  • A stronger reasoning model made the plan and split it into GitHub issues.
  • Each issue carried context, requirements, and acceptance criteria so another agent could pick it up independently.
  • Separate ChatGPT conversations (Sol 6 High) each got one issue.
• GitHub issues as the task queue
  • Workers read the project, edited files, ran tests, and iterated through MCP.
  • Sharing a filesystem required coordination, and a stronger model occasionally had to step in.
• Unlimited chats end, so the author moves to Claude
  • OpenAI announced it is ending unlimited chats and merging them with Codex usage.
  • The author says Claude Opus is smarter for their use cases and gives more mileage per subscription, so they are switching.

Community Reactions

• Criticism that people exploiting unlimited plans ruin it for everyone
  • It was a reaction to the author's claim that no rules were broken.
• Opinion that the limit change stems from OpenAI's funding troubles
  • The view is that OpenAI is becoming more capital-efficient.
  • The point is that it likely has nothing to do with this post.
• Experience that Claude's default limits are more generous
  • The commenter said using Sol for everything burned through limits.
• Opinion that multi-agent is mostly dead
  • The claim is that one big agent loop beats setups like this.
  • The commenter only uses sub-agents to compress context, such as tests needing 100+ tool calls.
  • They noted multi-agent schemes can bloat token usage by orders of magnitude.
• Experience that a $20 OpenAI sub with your preferred coding agent is a good deal
  • It came with the question of whether a Claude sub forces you to use Claude Code.

How to Apply

• Add acceptance criteria to your issue template
  • Split one task for AI into a GitHub issue today.
  • Fill in context, requirements, and acceptance criteria so another agent can pick it up independently.
• Separate planner and worker roles
  • Use the strongest model only for planning and issue breakdown on large features.
  • Give implementation to a cheaper model per issue, and reserve the strong one for design and review.
• Prevent file conflicts between parallel workers
  • State the file scope each issue may touch.
  • On a shared filesystem, use one branch per issue to coordinate.
• Build a tool-call log dashboard for MCP
  • Log file reads and command executions in your local MCP gateway.
  • Make failures inspectable so you can trace which calls go wrong.

Try It

# Paste into a Claude or ChatGPT chat (simplified version of the author's method: plan → issue breakdown)

You are a senior architect. Split the development task below into a list of GitHub issues that another AI agent can handle independently.

Task: {{development task description}}
Project structure: {{directories / tech stack}}

Each issue must include:
- Title
- Context (why it is needed)
- Requirements
- Acceptance criteria (in testable form)
- File scope to modify

Flag issues whose files overlap, and separate issues that can run in parallel from those that need ordering.

Next step: paste one issue at a time into a new chat and run them as workers.

Terminology

Related Papers

• GPT-5.6 Sol can beat Factorio on peaceful mode

  GPT-5.6 Sol in Codex at low thinking finished a peaceful Factorio run with a rocket launch

• Do the math: How we took over ChatGPT accounts

  One public share link ran attacker JS in a victim's ChatGPT session and could leak private chats

• AI isn't replacing SREs, it's deskilling them

  Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.

• Backdoor in Claude Code for VS Code extension on the Cursor Marketplace

  A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor

• Building an AI-native engineering org

  Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%

• We ran AI agent swarms in Minecraft and Mindustry

  We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.

• The Agent SDK

  Cursor released Agent SDK, letting you define agents in code and deploy them with one command.

• Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions

  Picking group-discussion memories by rarity or relevance often flips the final LLM decision

Related Resources

• https://chaosmanage.github.io/i-killed-unlimited-chatgpt/
• https://news.ycombinator.com/item?id=50319358
• https://news.ycombinator.com/item?id=50320255
• https://news.ycombinator.com/item?id=50312922