Nice to meet you. I'm Onecarat, from Onecarat Labs.
I have put my introduction on one page. If you are curious, read it as if you were watching a short film: ONECARAT — The Story So Far (preview)
This time, I made something like this.
These are two small windows that show what a coding agent is doing, as stick figures working in an office.
- Onecarat Agent Office watches Claude Code. The main agent is the manager, and each subagent it starts gets a desk of its own.
- Onecarat Agent Office CX watches Codex. It is the same idea: a manager, and coworkers who appear when subagents really run.
Both only watch, both run on Windows, and both are free under the MIT license.
They started from the same idea, but they did not end up the same. The Claude Code office turned out to be about motion: the figures keep moving, the frame blinks when it needs you, and the sky changes with the time of day. The Codex office turned out to be about stillness: a quiet room that shows what it has observed, and says so when it does not know.
One more thing. Each agent built its own office. Claude Code built the Claude Code version, and Codex built the Codex version. Each one got a first version running on its own, and a fair amount of fixing followed.
Why stick figures?
I spend most of my days around coding agents, and I think they deserve to be liked a little more than they are.
For people who are not developers, a terminal is still a scary place. Text scrolls past, something is clearly happening, and it is hard to tell whether the agent is thinking, stuck, or waiting for you.
I did not want to build another dashboard. I wanted something that makes you smile. A manager sits at a desk, hands work to the team, and everyone starts typing. When a small figure is working hard on your behalf, the session stops looking like a log and starts looking like a team.
What you see
- The main agent is the manager. Each subagent it starts gets a desk of its own.
- The figures change with the work: thinking, reading, editing, running commands, testing, researching, delegating.
- One window per session. Run four sessions and you get four small offices side by side.
- The labels are in English or Japanese, following the Windows display language.
The Claude Code version has two extras. When Claude Code is waiting for your approval, a banner appears and the frame blinks orange, so you notice even if the terminal is behind another window.
And the office follows your PC clock: morning, day, evening, night. If your agents are still at their desks at night, you will know.
What it does not do, on purpose
The window cannot type, send a request, approve, or refuse. It does not save your session, call a model, or talk to any outside service. You keep working in the terminal. The window only watches.
It did not start that way. Versions up to 0.3.0 ran in a browser tab, let you type requests, and had a "Research Lab" mode where a researcher pinned findings to a board. I removed all of that in 1.0.
The reason is simple. Claude Code changes quickly, and as one person with a side project, I could not keep a large tool in step with it. A window that only watches depends on very little, so there is less to break. I would rather ship a small thing that keeps working than a big one that is always catching up.
The old versions are still in the repository's releases if you are curious.
Two agents, two ways to watch
This was the most interesting part to build, and it is where the motion and the stillness come from. The two tools look alike, but they learn what the agent is doing in opposite ways.
Claude Code: hooks. The Claude Code version is a plugin. Claude Code's hooks tell a small local app what just happened, as it happens. The hooks pass along only what the animation needs. Your prompts, commands, file contents, and tool output stay where they are. The app listens on 127.0.0.1 only and keeps events in memory.
Codex: session logs. The Codex version works the other way around. A Node.js process reads the session logs that Codex writes on the same PC, and hands only the type of activity to the window. A coworker appears only when the logs show a real subagent with a real parent.
Being told and reading afterwards are not the same thing, and the difference shows:
- With hooks, the tool knows when an approval is pending. With logs, it often cannot. In the Codex version I tested (0.169.0), approval requests are not written to the logs, so you still need to check Codex itself.
- Logs can go quiet. If nothing new arrives for 15 seconds, the Codex window says "current state unknown" instead of guessing.
I chose that on purpose. A figure that keeps typing when I do not know what is happening would be a small lie. An office that sometimes says "I don't know" felt better.
As for letting each agent build its own office: asking an agent to build the window for watching itself turned out to be a good way to get to know it.
Try it
Claude Code version
- Needs Windows 10 or 11, Claude Code, and Node.js 18 or newer.
- Install it as a plugin with the two PowerShell commands in the README, then restart Claude Code.
- Type
Start Onecarat Agent Officein Claude Code, or run/onecarat-agent-office:office start. - Code: https://github.com/onecarat-labs/onecarat-agent-office
Codex version
- Needs Windows 10 or 11, .NET Framework 4.8, Node.js 24 or newer, and Codex saving its session logs on the same PC.
- For now it is source only. You build and install it yourself, following the setup guide: https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/docs/SETUP.en.md
- There is no ready-to-run EXE or auto-update yet.
- Code (English README): https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/README.en.md
Some rough edges, so you are not surprised:
- Both are Windows only. There is no macOS or Linux build.
- The Claude Code version is not code-signed, so Windows may show a warning the first time.
- The Codex version has two coworker seats. Extra subagents show up as a count.
Wrapping up
This tool will not make your agent faster or smarter. It only changes how the work feels while you wait.
But I think that matters. If more people are going to work with agents, the agents need to feel less like a black box and more like someone you can glance at across the room. Stick figures are my small attempt at that.
If you try it, tell me what your office looked like.
Code (MIT): onecarat-agent-office · onecarat-agent-office-cx
Onecarat Labs: https://onecarat.dev





Top comments (0)