Vibe Runner

Vibe Runner turns your real Codex workflow into a runner-first experience, letting you speak requests, make one-tap decisions, and hear results without stopping your run.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Vibe Runner

Vibe Runner

Inspiration

After a long day of vibe coding, I like to take a break and go for a run. The problem is that once I am out there, I start thinking about the project again—and I miss the vibe.

That idea began as a joke: if AI can keep building while I step away from my desk, why should going for a run end the session? But existing coding interfaces still expect sustained attention, typing, and screen management. None of them are designed around movement.

I built Vibe Runner to explore a different interaction model: speak the idea, glance only when judgment is required, tap once, and keep moving.

What Vibe Runner Does

Vibe Runner is a runner-first iPhone companion for the real Codex session running on your computer.

From a run, you can:

  • Speak a request instead of typing a prompt.
  • Continue working inside the actual repository on your computer.
  • Follow progress through glanceable status updates and audio.
  • Answer structured questions using large, one-tap cards.
  • Approve decisions without navigating a miniature desktop interface.
  • Hear the completed response while continuing to move.
  • Return to the same task and conversation afterward.

The goal is not to put an IDE on your phone. It is to make an AI coding workflow usable when your attention belongs somewhere else.

How I Built It

Vibe Runner includes a native SwiftUI iPhone app, a macOS desktop bridge, a native Windows bridge release candidate, and an encrypted relay.

The paired computer launches the real Codex App Server inside the selected repository. The phone and computer connect outbound through the relay, allowing the iPhone to continue the same task without exposing Codex or the host computer through a public listener.

Messages are protected with authenticated end-to-end encryption. The relay routes encrypted data but does not receive the message key.

The iPhone experience translates Codex events into movement-focused interactions:

  • Voice becomes the primary request mechanism.
  • Streaming activity becomes concise progress.
  • Structured questions become large decision cards.
  • Completed responses can be read aloud.
  • Projects and tasks remain connected to their real desktop counterparts.

I used Codex with GPT-5.6 Sol throughout Build Week to turn the initial idea into the SwiftUI application, desktop bridges, relay protocol, physical-device fixes, interface iterations, and regression coverage. I directed the product, design, and risk decisions while repeatedly testing the generated implementation against the real experience.

Challenges

The hardest challenge was reducing attention without pretending the product is completely hands-free. Runners still need to make decisions occasionally, so the interface had to make those moments brief, obvious, and safe: one glance, one large target, and one tap.

Another challenge was preserving the real Codex workflow. I did not want Vibe Runner to create a separate model conversation or become a remote-desktop skin. The task, repository, tools, and execution remain on the paired computer; the phone is a purpose-built control surface.

Remote connectivity also introduced security and reliability challenges. I needed to reconnect across networks without exposing a local service publicly, while ensuring the relay could not read private coding conversations.

Finally, physical-device testing mattered. Speech, audio playback, networking, app lifecycle behavior, and performance can behave differently on an actual iPhone than in the simulator. Several implementation choices changed only after testing the complete experience on hardware.

What I Learned

The biggest lesson was that powerful AI workflows do not always need more interface. Sometimes they need less.

Voice works well for expressing intent, but it is not the right answer for every interaction. Approvals and product choices are often faster as a glanceable visual decision. The strongest experience combines speech, audio, and one-tap controls rather than forcing everything through a single modality.

I also learned that agentic workflows become more useful when they adapt to the user’s context. The same Codex task can have a detailed desktop interface while seated and a dramatically simpler interface while moving.

Most importantly, the slightly ridiculous premise turned into something surprisingly practical. Ideas do not wait until we are sitting at a desk. Vibe Runner makes it possible to capture those ideas, keep the real workflow moving, and still let the run remain the primary activity.

Built With

Share this project:

Updates