Run your coding agents from your phone.

A software replacement for a physical command deck. Programmable keys launch Codex on a Windows host, each run in its own isolated worktree; progress, diffs and stop controls stream to a phone you can hold anywhere in the house.

See how a run works Local-first: no cloud, no API keys, no exposed ports.
Agent Deck · pairedShip keyset
PLANread-only workflow
BUILDisolated writable worktree
FIXisolated writable worktree
TESTisolated writable worktree
Run rack2 active
forge
hireos
quiz review ready
mocap review ready
sync stopped

How a run works

  1. Press a keySeven editable macros, two keysets, and a repository allowlist; nothing runs against a repo you didn't list.
  2. An isolated worktree opensOne fresh Git worktree and branch per run, so parallel agents never collide and nothing touches your working tree.
  3. Codex runs through its JSONL interfaceA small TypeScript bridge launches the locally authenticated CLI and translates structured events into readable progress.
  4. You watch from the phoneLive activity, the resulting diff, a stop control, and a resumable follow-up, with haptics and press-to-talk input.
  5. The branch survivesClean worktrees are archived safely and the Git branch is kept, so a run you liked is never lost to cleanup.

Built in

Six-digit pairing

One-time phone pairing backed by a stable local host token. The desktop makes the outbound connection; no port is exposed to the internet.

Durable run history

SQLite-backed runs and events, with restart reconciliation: a host reboot mid-run leaves a truthful record, not a dangling "in progress".

Honest status

Browser notifications on completed, failed and stopped. A disconnected companion reads as offline rather than quietly successful.

Scope, stated plainly

In this slice

  • Repository allowlist and macro definitions
  • One worktree and branch per run
  • Structured codex exec --json handling
  • Mobile activity, diff, stop, follow-up
  • Opt-in Windows logon task

Deliberately out

  • Live approval handling
  • Mid-turn redirection
  • Screen streaming and deployments
  • Cloud execution and API keys
  • Additional coding agents

The interesting part of agent work isn't typing the prompt; it's watching, judging and stopping.

See how a run works Local-first: no cloud, no API keys, no exposed ports.