Product

Platform

web

Electron desktop app (Chromium renderer) for macOS.

Users

A solo developer running several Claude Code agents in parallel across their own repositories, who wants to see at a glance what each agent is doing and step in (approve, answer, redirect) without hunting through terminal tabs.

Product Purpose

Run and steer coding agents from one place, shown as characters working in a pixel-art IT office. Each repo is an office floor; each agent is a character with its own git worktree and a real claude terminal. Success: the user can tell every agent's state in one glance and approve a permission request from the world in one click.

Positioning

Existing pixel visualizers (Pixel Agents, claude-office) only watch; existing orchestrators (Claude Squad, Conductor) are plain dashboards. This product owns the sessions it shows: it launches the agents, isolates them in worktrees, and answers their permission prompts from the world.

Operating Context

• The user signs in to their own unmodified claude CLI (their subscription or API key). The app never handles Claude credentials.
• Agents launched by the app get hooks through a per-session --settings file; nothing global in ~/.claude is changed.
• Worktrees live in ~/.agent-world/worktrees/<repo-slug>/<agent> on branch aw/<agent>. A quick fix can skip the worktree ("Work in my folder directly"), and non-git folders always work in place; several agents sharing a folder can overwrite each other. Removing an agent deletes its worktree only when nothing is uncommitted (or on an explicit discard), and its branch only when nothing on it is unmerged.
• Every agent is the user's own interactive claude or agy running in a terminal the app owns. The chat view is built on the program's saved transcript plus that terminal; the app never uses the Agent SDK. The agents' own work is never run headless. The app makes only short headless calls on the user's own account: claude -p for the usage limits, and, when workflows are on, claude -p or agy -p for the branch review and, at the Max dial, the intent check. Usage stays on the account's normal limits.
• Preferences (notifications, keep-awake, new-agent defaults, where claude is, update address) are kept in ~/.agent-world/preferences.json.
• Permission requests are decided in the world; after 5 minutes they fall back to Claude's own terminal dialog.
• On restart agents are asleep; clicking one resumes its session.

Capabilities and Constraints

• Claude Code and Antigravity (agy) are supported; more agent CLIs can be added behind one seam (docs/providers-design.md).
• Interfaces per agent: the office (state at a glance), Chat (conversation, tool cards, pictures, ! shell mode, permission modes, rewind, /btw side questions, Ctrl+R history, queued messages), Terminal, Code (files and diffs), Tools (MCP servers and plugins: add, remove, enable), Preview (a dev server the agent started) and History (resume an earlier session).
• Anything Claude asks the user (tool approval, plan, folder trust, questions, its own menus) appears in an 8-bit dialog or inline in the chat, and comes back after a window reload.
• Agents are hired from roles (built-in, saved, or a project's .claude/agents), can be grouped, cloned, and handed to another agent for review; commands started by an agent can be listed and stopped from the chat.
• Usage limits (5-hour, weekly) are shown as the tanks of the power plant on the back lawn of the office; the top bar's Usage button opens the full picture and counts the limits running low.
• Shipping: installers for macOS through electron-builder (signed ad hoc until a Developer ID certificate is provided), update checks against the project's GitHub releases or an address set in Settings, and a CI workflow on macOS. There are no Windows or Linux installers.
• Roles are dynamic: an agent is a senior; while it has live subagents it is the boss; subagents are interns.
• Activities shown: idle, thinking, researching, implementing, running, delegating, waiting, permission, asleep.
• No external art assets in v1: all pixel art is generated in code (no downloads, no license questions). Cosmetic upgrades and themes come later.
• Undecided: product brand name (code name agent-world); the name must not include "Claude" or "Anthropic".

Brand Commitments

• Theme: IT / engineering office.
• Visual reference supplied by the user: a top-down 16px pixel office with a wood-plank workroom, a back wall with bookshelves, a clock and hanging plants, CRT desks, a meeting table, a blue-floored lounge with sofas and framed pictures, and a checker-floor kitchen, with dark wall outlines between rooms.

Evidence on Hand

No real users, testimonials or benchmarks exist; none may be invented.

Product Principles

1. Control over spectacle: every visual state maps to a real agent state the user can act on.
2. Own the session, don't scrape it: prefer launching and hooking agents over reading local files.
3. Never lose work: uncommitted changes block removal; nothing is deleted silently.
4. Nothing global, nothing hidden: no changes to the user's Claude config, no credentials touched.

Accessibility & Inclusion

Every world action (select an agent, approve, deny, wake, remove) must also be reachable from the side panel with a keyboard.

Edit this page on GitHub