Alondra

Run a team of coding agents on your Mac and steer them from your iPhone.

Alondra is a daemon that runs Claude Code and Codex sessions on your Mac, turns each piece of work into a task with its own worktree and pull request, and brings every decision to you as a card: a question, a plan, a permission, a finished pull request. You answer from the terminal, the Mac app or the iPhone; the agents carry on.

Your agents run on your Mac; Alondra Cloud connects your devices. Teams are coming soon.

The terminal, Mac and iPhone apps on a demo sandbox with made-up projects.

Install

brew install bitomule/tap/alondra
alondra

alondra sets up the rest the first time: your projects folder, the background service and the Claude Code plugin. Run it again at any time; it only changes what is missing.

You need a Mac on Apple Silicon with macOS 26 or later, and Claude Code, Codex or both, already installed and signed in. INSTALL.md has the details, the iPhone and Mac apps, and what to do when something fails.

Before installing, read what Alondra does on your Mac: installed programs, files, connections, permissions and credentials, with source links and a prompt you can give an agent to audit the checkout.

What it does

1. You write a request in the TUI, the Mac app or the iPhone: "add rate limiting to the API".
2. The orchestrator reads it, makes a plan and splits it into tasks. A plan that needs your approval comes back to you as a card for approval before any code changes.
3. Each task gets a node: an agent session in its own git worktree, running Claude Code or Codex with your own account. The orchestrator picks the model and the reasoning effort of each task, and Alondra runs it on one of your accounts. A task that depends on another waits until that one has landed; the rest work in parallel, within what your Mac and your plans allow.
4. Cards bring you in only when needed. A node that has a question, wants a plan approved or opens a pull request puts one card in front of you. Nothing else interrupts you.
5. Alondra merges and cleans up. When you accept a delivery it merges the pull request and removes the task's worktree.

By design, every task and node has one visible state, and a task waits only for capacity or for you. A task that gets stuck anyway is a bug, and fixing those is part of the work.

How it works

Piece	What it is
alondrad	The daemon: runs 24/7 under launchd, keeps the state in SQLite and serves the API
Orchestrator	Turns requests into plans and tasks; runs only when there is work, never as an idle process
Nodes	One agent session per task, in its own worktree, through the Claude Agent SDK host or Codex
Cards	Questions, plans, permissions and deliveries, each decided from the card alone
alondra	The CLI and TUI, and the MCP server the nodes use to ask you things
iPhone and Mac apps	The same cards and tasks, with push notifications

Only sessions that Alondra launches take part: a Claude session you start in your own terminal has no plugin, no task and no card. The design, from the daemon's state model to the read side the screens use, is in docs/architecture/overview.md.

This is the beginning of the README. Read the rest at the original. Read the full README ↗