Inspiration
From my experience using Codex, I once had only 20% of my usage limit remaining while there was still a lot of project work to complete. Sometimes, Codex performed tasks outside my instructions, so I needed to make the most of my remaining limit.
That experience inspired me to create RX Context Guard, a protective guardrail designed to keep AI agents working within the instructions they are given and prevent them from performing unauthorized tasks.
However, RX Context Guard is still flexible. When an AI agent proposes an additional action, the developer can review the proposal and decide whether to ALLOW or REJECT it. This way, the developer remains in control.
What it does
RX Context Guard checks whether a proposed AI action is allowed under the defined rules. If the action falls outside those rules, the developer can review it and decide whether to ALLOW or REJECT it.
How we built it
We built RX Context Guard using Next.js and TypeScript. We started with an AI action simulator because I wanted to test whether RX Context Guard worked correctly before connecting it to a real AI coding agent. The simulator allowed us to test the Guard Engine, action boundaries, developer approval, and rejection evidence.
Challenges we ran into
The biggest challenge I faced was verifying that an AI-proposed action was truly not executed after being rejected. I wanted more than just a REJECTED status. I needed evidence that the rejected action had not been performed. To address this, we added rejection evidence showing performed = false in our simulation, making the result easier to verify.
Accomplishments that we're proud of
The most memorable achievement for me was when we rejected an AI-proposed action in RX Context Guard. The system displayed REJECTED, and the rejected action was not executed in our simulation, as confirmed by the execution evidence showing performed = false.
What we learned
Sometimes, AI suggests good ideas that we had not thought of before. That is why RX Context Guard is designed to be flexible. We can review AI proposals before making a decision. If a proposal is useful, we can approve it. Otherwise, we can reject it. This allows us to benefit from AI suggestions while keeping control in the developer's hands.
What's next for RX Context Guard
My next step is to integrate RX Context Guard directly with Codex and other AI coding agents, so its guardrails, review process, and developer approval controls can be applied to real coding tasks, not just simulations.
Log in or sign up for Devpost to join the conversation.