TryBrake for ChatGPT & OpenAI
OpenAI agents that ask before they act.
Whether it's Codex writing code in your terminal or an agent you built on the OpenAI API, TryBrake checks each risky action against your rules first: allow it, block it, or hold it for a human.
Codex hooks cover shell commands. For your own code, any OpenAI model works.
How it works
Set up in three steps.
- 1
Pick your setup
Codex CLI, or your own agent in Python or JavaScript (OpenAI Agents SDK, Assistants, function calling).
- 2
Paste the install text
Paste it into Codex or into your AI coding assistant. It adds TryBrake's hook, or the connector file and your key, to the project.
- 3
Wrap the risky tools
For your own agent, one line per tool: guard("stripe.refund"). The tool then only runs if your rules say yes.
What it stops
The mistakes you can't undo.
Codex shell commands
rm -rf, database wipes and curl | sh are blocked; force pushes wait for you.
Function calls
Each tool your OpenAI agent can call gets its own rules, using the real arguments, like the amount of a refund.
Human approval
Hold an action until someone approves it from the dashboard, Slack, Teams, Google Chat or Telegram.
Replay
See every step of a run with the agent's reasoning, what was allowed, blocked or approved.
Questions
Good to know.
Does it work inside the ChatGPT app itself?
Not inside the chat window. TryBrake protects agents that take real actions: Codex, OpenAI's coding agent included with ChatGPT plans, and agents you build with the OpenAI API.
Do you need my OpenAI key?
No. Your agent keeps its own keys. TryBrake only answers yes, no or wait.
What if TryBrake is unreachable?
The action doesn't run. Every connector fails closed.
Works with more than ChatGPT & OpenAI.
One TryBrake account protects every agent you run, wherever it's built.
- n8nProtect workflows and AI Agent tools
- MakeGuard modules in your scenarios
- Claude CodeBlock risky commands and edits
- ZapierOne webhook step before the action
- CursorTerminal commands and MCP tool calls
- Gemini CLICommands and file edits
- PythonLangChain, CrewAI, OpenAI Agents SDK…
- JavaScript / TypeScriptVercel AI SDK, LangChain.js, Node
- Anything elseOne HTTP call before each action