Why we built Printhouse
AI agents are powerful. The tools around them aren't. We built the workspace that was missing.
There’s a weird gap right now between what AI agents can do and what the tools around them let them do.
The models are incredible. They can write code, debug systems, analyze data, compose emails, build entire applications from a description. The raw capability is there. But the moment you try to actually use an AI agent for real work, you hit walls everywhere.
Your laptop has to stay open. You’re approving permission prompts every 30 seconds. The agent can’t install the package it needs. It forgets who you are between sessions. It can see your files but can’t actually run code. It has “tool access” but the tools are toy wrappers around limited APIs.
The experience feels like giving someone a fully stocked workshop and then handcuffing them to a chair.
What we wanted
We wanted something simple: give the AI a real computer and give us a real workspace to collaborate with it.
Not a chat window with some plugins bolted on. Not a terminal-only interface for power users. A proper workspace — like an IDE meets Slack meets a cloud dev environment — where the AI agent has genuine, unrestricted access to a real machine and you can see everything it’s doing.
We wanted the agent to run while we slept. We wanted it to connect to our tools without us handing over raw API keys. We wanted to organize our work into channels so the same AI could help with engineering in one context and career planning in another. We wanted it to get better over time instead of starting from scratch every conversation.
What we built
Printhouse is the workspace that came out of that. Each user gets their own persistent cloud computer running a full Linux environment. The AI agent lives on that computer. You interact with it through a collaborative workspace — chat, file explorer, terminal, code editor — all pointing at the same machine.
The agent installs packages, runs servers, writes scripts, manages files, and takes action across your connected tools. It doesn’t ask permission for routine operations. It doesn’t lose context when you close your browser. It doesn’t pretend to do things — it actually does them, because it has a real computer.
A few architectural decisions we’re proud of:
The credential proxy. Your API keys never touch the agent. A sidecar proxy intercepts outbound requests and swaps dummy placeholder keys for real credentials at the network level. The agent literally cannot see your actual API keys. This was non-negotiable — if you’re giving an AI agent access to your Gmail and GitHub, the security model has to be real.
Channels. Your workspace is organized like Slack. Each channel has its own instructions, context, and tone. #eng gets technical instructions about your codebase. #career gets context about your job search. Same AI, different modes. This turned out to be more powerful than we expected — it’s the natural way to work with an AI that touches multiple areas of your life.
Skills. Reusable instruction sets that extend what the agent can do. They’re just markdown files with structured instructions, invoked via slash commands. A skill might define a full code review pipeline, a marketing copy framework, or an SEO audit procedure. There’s a community marketplace for sharing them.
Where we are
Printhouse is in research preview. The product works — we use it every day, and we literally built it using itself. But it’s early, there are rough edges, and we’re adding people gradually to make sure the experience is solid.
If this sounds like the thing you’ve been looking for, we’d love to have you try it.