My first open-source project: a pixel-art office where my coding agents work, and I'd love for you to try it

Agent World — Tue, 27 Oct 2026 16:32:13 +0000

Agent World runs and steers your coding agents from one place, drawn as characters working in a pixel-art office.
Code, guide and downloads: https://github.com/hodkep/agent-world (Apache 2.0)
For Mac for now. Windows and Linux are coming soon.

This is the first open-source project I have ever released. I am proud of it, a little nervous about it, and I believe it is good enough that you should try it and tell me what you think.

How it started

Coding agents changed the way I work. I hand a task to one agent, then another to a second, then a third, and for a while it felt like having a small team.

I tried a lot of tools to manage that team: grid views, terminal-based agent managers, polished interfaces, even Claude's own app. Each of them was fine, and each of them bored me within a day. They all showed me the same thing, terminals in boxes, and a wall of boxes is still a wall of boxes. I wanted something I would actually enjoy opening every morning, something built the way I like to work, that I would keep using on my own long after the novelty wore off.

So I built it. An office I can look at.

Every agent is a character at a desk, and what the character does is what the agent does: thinking, researching, implementing, running a command, delegating, or waiting for you. An agent that is asleep leaves the floor but stays in the list. One glance tells me where I am needed.

I wanted it to be fun to look at, but I did not want it to be a toy. Three rules shaped everything:

Control over spectacle. Every visual state maps to a real agent state that you can act on.
Own the session, do not scrape it. The app launches and hooks the agents it shows. Your own claude stays unmodified, and nothing global in ~/.claude is changed.
Never lose work. Uncommitted changes block removal, and nothing is deleted silently.

It works with your own claude (Claude Code) or agy (Antigravity) sessions, and it never touches your login.

Then I kept adding what I wanted

Once the office worked, I kept going, one thing at a time, the way you do when you are building a tool for yourself.

More insight. I wanted to know where the tokens and the time went, which agent was doing what, and what each run had actually followed, so I built views for all of it.
More autonomy, and more determinism. I wanted agents that could carry a task further on their own without wandering off, so there are workflows, a playbook of rules, and a push that waits until I have seen the findings.
Everything in one place. A web preview of what the agent is building, its files and code, a terminal, and the option to select anything, words in a reply, lines of code, terminal output, and add it as context for the next question, all inside the app, without copying and pasting between windows.

What makes it different

Watch several agents at once. Grid mode puts agents side by side, and each pane shows either the agent's chat or its real terminal, whichever you need. Read one conversation while another agent runs its tests, without switching away from either.

Answer an agent without leaving the room. When an agent needs permission, the request appears right above its character. Allow or deny it in one click, or answer in the agent's own terminal. The agent number beside the chat counts who is waiting for you, so nothing sits blocked while you look elsewhere.
Workflows that put you in charge. A workflow is a plan you approve, checks that run, a review of the branch against the plan, and then the pull request. Each run is drawn as a flow, one card per step, and you can see where it stands and which step is waiting for you.
A push that waits until you say so. Every agent works on its own branch in its own git worktree, so five agents never trample each other. With workflows on, git itself holds an agent's push until the plan, the review and its findings are settled. You release it from the agent's chat.
A playbook for your whole team. Keep skills, instructions and prompts once, for every agent and every program, and write down which agent is told what, and when. The chat shows what the library did for each message, so the rules are never a black box.
One board for you and your agents. Lists, tasks that need you, tasks an agent has claimed or finished. A task has subtasks as a checklist with a progress bar, and labels you add, rename and recolour. When an agent writes `todo#14` in a reply, that is a link to the task.
Agents that draw and ask properly. Agents draw diagrams in their replies and ask with real forms. Click a box to pick it and ask about it, or open the diagram large to zoom and draw on it. A wall of text becomes something you can point at.
Claude Code and Antigravity, one team. Hire agents on `claude` or `agy` side by side and let them message each other. A Claude agent can write to any other agent right away, and an agy agent joins in after one click on Connect in Settings. One shelf of skills and one playbook serve both programs.
Several Claude accounts at once. Sign in more than one subscription, hire each agent on the account you choose, and move agents between accounts without signing in again. Or switch every agent of a project to another account in one go, and they continue where they were.
Switch agents with one click. The number beside the chat opens a panel with a tile for every agent and what it is doing. Click a tile to jump to that agent, or step through them with Ctrl+Tab. Projects work the same way, with a red number on the one that has an agent waiting for you.

And more

The README and the guide cover the rest, so here is just the list:

Exact token counts by agent, project and model, and a Weekly Rewind
A Notes window that stays above every other window, where a checklist line can become a task
Side questions with /btw that do not interrupt the agent
A terminals drawer for each project, and a shortcuts sheet on Cmd+/

Please try it

brew install --cask hodkep/tap/agent-world

Or take the dmg from the releases page, arm64 for Apple silicon and x64 for Intel. The app is not signed with an Apple Developer ID yet, so the first launch needs System Settings, Privacy & Security, then Open Anyway. The Homebrew cask skips that step. Nothing else has to be installed first, because the app carries its own git and helps you set up the Claude Code CLI.

It is young, it has rough edges, and for now it is for Mac only. Windows and Linux are coming soon. I would genuinely like to hear:

Did setup work the first time? Where did you get stuck?
What would make you reach for it every day?
What is missing?

Let's build your ideal development tool

I started this because I wanted a better way to work with coding agents, and I am quite sure my idea of "better" is not the only one. So here is my ask: tell me what you would want in your ideal development tool, and let us build it together.

Bring a half-formed idea, a pet peeve from your own agent workflow, a feature you wish existed, or a rough prototype. Open an issue or a Discussion, or send a pull request. CONTRIBUTING.md explains how to build it and run the tests, and small contributions are as welcome as big ones.

If you like it, a star on GitHub means a great deal to a first-time maintainer. And if you do not, tell me why. That helps even more.

Code, guide and downloads: https://github.com/hodkep/agent-world

More good ones coming.