We built an office view for Copilot, Claude Code and Codex

Vikram Gorla — Thu, 12 Nov 2026 21:34:32 +0000

I was running many agent sessions at once. Copilot in one window, Claude Code in another, Codex in a third, and often more than one folder for each. I kept switching between those windows to see who was still working and who was waiting for me.

After a while I was not doing my own work. I was watching the agents work. I had wanted to multitask, but the switching became the job. It was stressful, and it was not productive.

We built Agent Durbar to put those sessions in one view.

What it is

Agent Durbar is a desktop app for your digital agents. It is an office view. Each folder becomes a room, and each session has a desk. From one screen you can see who is working, who is thinking, who is waiting for you, and who has run into trouble.

A room is any folder an agent works in, not only a code project.

It works with the Copilot, Claude Code and Codex apps already installed on your machine. Durbar is not a new model provider. It drives those apps locally, and usage still goes against your existing plans.

What you can do

Click a desk to open a full-screen chat with that session. You see the agent's thinking, tool calls and output as they come in, including diffs when a file is being changed. You can attach files and images by picking them, dropping them or pasting.

If you have VS Code, VS Code Insiders or VSCodium installed, you can open the same folder from the chat header. The conversation stays in Durbar.

When an agent starts helpers (sub-agents), those helpers sit under the agent that started them. When the office has room, you can see what each one is doing.

About 10 to 15 agents stay readable on one screen. We checked this with 15 real Copilot sessions at once, on Linux. When there are fewer agents, the desks grow and show more of the thinking or the last reply.

A "Needs you" inbox collects every session that is waiting for an answer or a permission, oldest first. You do not have to click through every desk to find the one that needs you.

If a session should be handed on, you can start a new session from a hand-off summary of the current conversation, and you can edit that summary before it is sent. If you have more than one agent app, you can also continue a conversation in another app, in the same folder.

Sessions resume after a restart. There are also shared MCP server settings and scheduled prompts, if you need them.

How we built it

Most of the app was built by Claude Code agents that our team directed and reviewed, working from written specs.

The README has a build cost page with the numbers. They are API list-price estimates, not billing. As of the current release, that estimate is about US$5,398.

Honest limits

The current stable release is v0.1.17. Builds are available for Linux, macOS and Windows, but they are unsigned and not notarized, so your operating system will warn you the first time. There is no auto-update.

Linux is used every day. macOS has been tried by hand on two Apple Silicon Macs. Windows has had one hands-on look on a real PC. CI builds and checks all three platforms. The per-session spending limit is a soft limit. The README links to the known limits.

Durbar runs locally. It has no accounts, servers or telemetry, and it does not see your tokens. The agent apps handle their own sign-ins. New sessions start in Read-only. Broader access is a deliberate choice.

How to try it

Agent Durbar is free and open source, under the MIT license.

Website: three-feathers.github.io/agent-durbar
Downloads: latest release
Source: github.com/three-feathers/agent-durbar

There is a short demo video if you want to see the office view first.

If you want to look at the office without using any credits, you can start a local clone with the built-in simulator (DURBAR_SIM=15 npm start). It shows 15 agents and starts no real sessions.