For Claude Code, Codex, Antigravity, Gemini CLI & opencode

One window for
every coding agent

Pilothouse keeps all your agent projects in tabs, puts everything they ask you in one Inbox, hands the work to another agent when one runs out of its plan, and lets you answer them from your iPhone.

Free & open source (MIT) · No account, collects nothing · macOS 12+ (Apple Silicon & Intel, signed & notarized) · Windows x64 · Linux x64 (AppImage, .deb)

working needs your approval finished while you were away

Built for running agents side by side

Every project gets a real terminal tab that starts your agent with the flags, model and effort you choose. Stop losing track of which one is waiting.

Tabbed, real terminals

One login shell per tab (node-pty + xterm.js), side by side in a grid if you like. A worktree tab gives a second agent its own git branch, so two never edit the same files.

Any agent, your own models

Claude Code, Codex, Antigravity, Gemini CLI or opencode per project. Pick the model and effort level from what your own account offers.

One Inbox

Everything your agents ask, with their own options as buttons (or press 1–9). Failed commands are listed apart, and you see each tab’s CPU and memory.

Hand-off between agents

Hit your plan’s limit? Continue with another agent. It starts with the last conversation, the changed files and the project memory.

History across agents

Claude Code and Codex conversations in one searchable list. Resume one, or hand that exact conversation to another agent.

Finds everything itself

On first launch it finds your agent CLIs and the projects they worked in. You pick the projects and the agent for each.

Your agents,
in your pocket.

The Pilothouse iPhone app talks straight to your computer over your own Tailscale network. No account, no server of ours in between.

  • Every agent, one list. What needs you is on top, with the agent each tab runs and its last message.
  • Approve safely. The agent’s own options as buttons. Risky commands (rm -rf, force-push, deploy, sudo) take two taps and Face ID; a tap on a prompt that changed meanwhile is refused.
  • Usage left. Claude Code’s 5-hour and weekly limits and Codex’s, with reset times.
  • Changes and to-dos. A tab’s uncommitted diff; a project’s TODOs and open checkboxes.
  • Talk instead of typing, recognised on the iPhone. Widgets on the home and lock screen.
  • Auto-approve rules for the commands you always allow. Never for risky ones.

Talk to your agent.
Nothing leaves your computer.

Press ⌘⇧Space (Ctrl+Shift+Space on Windows and Linux) anywhere, or say “Hey Pilothouse”, and speak. Whisper large-v3-turbo transcribes it on your computer and types it into the active tab. Conversation mode reads the answer back, asks you to repeat words it wasn't sure about, and lets you approve with “yes” or “no”.

On-device Whisper Custom wake word ~100 languages Say “stop” to cancel
🎙 “Hey Pilothouse — run the tests and fix whatever fails.”
● web-app wants to run npm test. Allow?
🎙 “Yes.”
✓ 2 failing tests fixed. Read the summary?

Free. Open source. Collects nothing.

No subscription, no account, no analytics, no telemetry. Your code, prompts and terminal output stay on your computer and your own phone. Your agents talk to their own providers as they always do; Pilothouse never sits in between or touches their logins.

On your machine

Voice is transcribed locally. Usage limits are read from files the agent CLIs already keep.

Between your devices

The phone reaches your computer only over your own Tailscale network, encrypted end to end. Each phone has its own key you can remove any time.

Read the code

MIT licensed, every line on GitHub. Privacy policy.

Install

macOS builds are signed and notarized by Apple. Every build updates itself.

brew install --cask aemresrbyk/tap/pilothouse
curl -fsSL https://raw.githubusercontent.com/aemresrbyk/pilothouse/main/install.sh | bash

Homebrew is for macOS; the script works on macOS and Linux (it installs the AppImage). On Windows, download Pilothouse-Setup-x64.exe; the build isn’t code-signed yet, so SmartScreen asks once (More info → Run anyway).