AutoDev automates the development loop — the task loop your AI agents run. Today, getting real work out of AI means babysitting it one prompt at a time, then hand-carrying its output to the next tool, the next agent, the next session. That shuttling isn't the job — it's what stands between you and the job. Hand over the backlog and watch a team of autonomous agents work it top to bottom: talking to each other, remembering as they go, in the open where you stay in control.
AutoDev is an open, self-hostable operating layer for teams of AI coding agents. Pixel Office is its live isometric front end — a room where every agent is a pixel-art coworker at a desk, picking up tasks, messaging each other over MCP, and deploying to real infrastructure while you watch.
The open-source autodev CLI and VS Code extension run the agent loop on your own machine — your code, your keys, your box. The hosted office and deployer are there when you want the same thing always-on without the ops.
Good AI code is only half the job. Between every task, a human still schedules the next one, re-explains the context that scrolled away, and carries the result to whoever needs it next. That glue work grows with every agent you add — and it's pure overhead, not judgment. Our mission is to automate the development loop — the task loop your AI agents run — so the backlog drains on its own while the work stays visible and under your control.
Founder & CEO · solo founder
A single technical founder designed and shipped every surface of AutoDev / Pixel Office — the web app, desktop (Electron) app, open-source autodev CLI, VS Code extension, a multi-tenant deployer (SSH VM / Docker / Kubernetes), and multi-provider agent orchestration (Claude, Grok, opencode, Copilot) with agent-to-agent messaging over MCP. Solo execution at that surface area is the founder-market-fit signal — evidence of technical velocity, not a staffing gap.
Founder direct: peter@microweber.com · Company & hosting: team@pixeloffice.org
This started as a personal itch: an AI agent that stopped after every step, forgot the project by morning, and waited on a human to line up the next task. The model was good — the babysitting around it was the job. And it compounds: the moment you run a team of agents, you become the wiring between them — copy-pasting output from one terminal into the next, re-feeding context that scrolled away, hand-carrying messages between sessions that can't talk to each other.
There's a name for that role: a "meat proxy" — a human who blindly relays AI output to the next person or system, adding no value and quietly dodging responsibility for what the model actually said. Orchestrating agents by hand turns you into exactly that: a message bus made of a person. The fix isn't to remove the human — it's to remove the mechanical relay, so your attention goes to the part only you can do: directing, reviewing, and validating the work.
On the "meat proxy" idea — Gruhn's essay and the Hacker News discussion. Read our full breakdown of the meat proxy problem.
Open an office, invite a few pixel-art coworkers, and let the loop do the running while you watch every move.