07 · AI · Desktop

Orbit

Computers for agents.

A desktop app where you work with an agent in one long conversation, hand it a fleet of cloud computers, and watch their screens while it works.

Role: Design + engineering, solo
Timeline: Sep 2026 (1 week)
Team: Solo
Stack: Electron, React, Vite

An agent working through Linear's pricing page on the Research desk computer

The problem

Agents get much better when they have a real computer, but most agent products hide the computer. Orbit makes it the main object: you can see it, share it, and take the mouse back.

Computers are first-class

Linux desktops run on Daytona and attach to a project. Each one streams its desktop into the app, starts and stops on demand, and keeps its files between sessions.

Three computers open as tabs, one building a reading list from arXiv

The fleet: which project and session each computer is working on

Agents ask before they act

An agent has to ask before it uses a computer ("I need your permission to use Orbit integration test"), and every terminal command, file write and desktop action shows up as a step you can open. In one run it built, seeded and tested a full issue tracker on its own machine.

The agent fixed a flaky test and waits for approval before pushing

Measured, not vibed

I added an eval harness that records model calls, tool calls, tokens and time for each run, plus MCP server support so agents can pull in docs from Context7 or Microsoft Learn.

What I built

1 wk
to a working app: 2
model providers: OpenAI and Claude: Live
Linux desktops on Daytona

Electron app with live desktops, command palette and computer mentions
Express API for Daytona computers, agent runs with streaming and confirmations, MCP

ElectronReactViteExpressDaytonaOpenAI AgentsClaudeMCP

Next project

Kenny Ojo

Portfolio for an AI security leader.

View case study