# Agentbox

> An open-source Mac app that gives coding agents an inbox, running Claude Code and Codex on your own subscription.

Agentbox is a free, open-source app that turns coding agents into rows in an inbox, so you hear only from the ones that are done or stuck. It runs Claude Code and Codex locally on your own subscription, with no server and no account. It is available as a Mac app or in the browser via npx, under GPL-3.0.

## What It Is

Agentbox replaces the chat window with an inbox or board of agent threads. You paste a list of tasks, send over issues from a board, or type a single task. Each item becomes its own row and its own agent, already working. Each row summarizes what the agent did and what it needs from you, often with a recommended option you can accept with a keypress.

## How the Workflow Runs

Agents work in the background and a row returns to you only when its agent finishes or gets stuck. You can approve a recommendation, pick between options the agent offers (including design variants drawn in your actual app), or type a reply that sends the work back. Code changes, designs and docs can be reviewed in the app. According to the site, each task gets its own git worktree and a built-in terminal, and tasks can be snoozed or set to repeat on a schedule. Desktop notifications fire when threads need you.

## Priorities and Keyboard Control

Tasks can be tagged Urgent, High, Medium or Low, and projects can be ranked by dragging. When tasks outnumber available session slots, the queue starts the most important work first. Only your own tags count; an agent cannot promote its own work. The app is keyboard-driven, with J/K navigation, number keys for options, and a Command K palette for everything else. The README also lists fifteen themes in light and dark.

## Setup and Privacy

Both the browser and Mac app routes need Node 22 and a signed-in Claude Code or Codex. Run `npx agentbox-app` for the browser version, or clone the repo and run `npm start` for the Mac app. By default, spawned sessions are granted only the store tools; letting agents edit code and run commands is an explicit opt-in in the config. The README says builds from the repo send no diagnostics, while signed released builds can send limited counts, which can be turned off in Settings.

## Features
- Inbox and board views of agent threads
- One agent per task, each with its own git worktree
- Project ranking and Urgent/High/Medium/Low priority queue
- Command K palette and keyboard shortcuts
- Built-in terminal and diff review
- Review of code, designs and docs in the app
- Recurring scheduled tasks
- Snooze and later reminders
- Desktop notifications
- Design options drawn in your own app
- Send-back replies to agents
- Fifteen light and dark themes
- Imports existing Claude Code agents
- Configurable concurrent session limit

## Integrations
Claude Code, Codex

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Links
- Website: https://agentbox.ac
- Documentation: https://github.com/savannahfeder/agentbox#readme
- Repository: https://github.com/savannahfeder/agentbox
- EveryDev.ai: https://www.everydev.ai/tools/agentbox