What it does

AI work creates an unusual gap: the agent may be thinking, testing, or editing, but the person keeps staring at the screen. ReApp turns that handoff into a short, user-controlled breathing reset.

Inside ChatGPT and Codex, someone can ask for a 30, 60, or 90-second pause in plain language. ReApp returns a visual pacer with a fixed five-second inhale and five-second exhale, pause/resume controls, reduced-motion support, and an immediate early exit. GPT-5.6 can tailor a small amount of cue language from bounded, non-sensitive inputs. If that request is unavailable, reviewed local copy keeps the experience working.

The optional Mac companion lives in the menu bar. It can schedule a full-screen pause at a randomized interval between 30 and 90 minutes, offers one-click duration controls, and exits with Escape. It runs independently and does not read Codex prompts, ChatGPT conversations, source code, or screen contents.

Inspiration

Long stretches of screen work can make it easy to stay physically still and mentally locked onto the next result. AI makes that pattern stranger: we hand work away, then watch the machine work anyway. ReApp began with a simple question: what if the handoff became permission to step back for less than two minutes?

How we built it

ReApp is an MCP-backed app plus a bundled Codex skill. Three typed tools plan the reset, render the interactive widget, and return a completion cue. The Responses API uses GPT-5.6 with Structured Outputs to keep personalized copy inside a small schema. Fixed application code—not the model—controls allowed durations, five-in/five-out pacing, cycle counts, input length, and fallback behavior.

The native companion is a separate SwiftUI Mac app distributed as a signed and notarized DMG. The public Node service and landing page deploy on Vercel, while the Mac release is hosted on GitHub. The original ReApp iOS repository remains unchanged; this AI companion was built in its own repository during Build Week.

Codex helped architect, implement, test, and ship the MCP server, SwiftUI companion, Vercel deployment, signing and notarization pipeline, and submission package. I directed the product boundary, breathing cadence, user-control requirements, copy, and the decision to leave the original iOS app unchanged. GPT-5.6 powers only bounded cue personalization; deterministic application code owns timing, pacing, validation, and fallback behavior.

Challenges we ran into

The hard part was not drawing a breathing circle. It was making the pause feel trustworthy. We kept the menu-bar controls readable, eliminated a distracting full-screen layout transition, made Escape immediate, prevented broad “reset” language from triggering the wrong tool, and separated a real host capability from a marketing promise. ReApp can accompany independently running work, but it cannot pretend that every ChatGPT or Codex task continues in the background.

Accomplishments we are proud of

One concept works as a Mac menu-bar companion, responsive browser reset, MCP app, and Codex skill. The timer stays functional without an API key or model response. The tool contract is stateless, bounded, and explicit about sensitive data. The native beta is signed, notarized, and distributed separately from the untouched iOS app. The interface supports reduced motion, pause/resume, early exit, and 30/60/90-second choices.

What we learned

A humane AI feature needs stronger boundaries than a clever prompt. The most important design decisions live in deterministic code: when the tool can trigger, what it may receive, what pacing is allowed, what it must never claim, and how quickly the user can leave. The model contributes tone; the product contract protects agency.

What's next

Next comes a small public beta, OpenAI plugin review, and feedback on whether the handoff moment is genuinely useful. Future work may include more local-only preferences and additional accessible cue styles. ReApp will not add stress detection, medical scoring, or hidden monitoring.

Built With

  • chatgpt-apps-sdk
  • codex
  • gpt-5.6
  • model-context-protocol
  • node.js
  • openai-responses-api
  • swiftui
  • vercel
Share this project:

Updates