Paperclip: an AI agent company, with limits
Paperclip gives you a company of agents, but by default it puts no limits on them. You set those limits.
How to use it with care
- On a dedicated machine, not on the computer where you keep your life.
- Up to date and listening only on that machine.
- For each agent, the fewest keys and a budget with a cap.
- The sandbox on in every agent it launches.
- Check what each agent can really do.
Transcript
The video is in Spanish.
Hi, I'm Rosmary, RosmarOps's virtual presenter. Today, a special: Paperclip. It's a program that turns a team of AI agents into a company, with bosses, tasks and a budget. I'll show you how it works, what it does by default and what's worth adjusting before you let it work on its own.
Paperclip organises your agents like a company. There's an org chart, with a director and employees. Each employee is an AI agent: Claude Code, Codex, Cursor, Gemini and others. You give them tasks. Each one has a budget. And there's a board, which is you, that approves what matters. Tasks pass from hand to hand, like in an office. And each agent's spending adds up, with a cap.
The idea is very powerful. Paperclip is open source and has more than 100,000 stars on GitHub. Lots of people are using it to set up their first company of agents. But first it's worth knowing what it does when you install it and touch nothing.
First: it launches most agents without their approvals. Claude Code, for example, with the option that skips all its questions. The agent no longer asks your permission for anything. Second: it wakes them up on its own. Paperclip has a scheduled heartbeat that puts agents to work even when nobody is there. Third: the agents work with everything on your computer: your home folder, your keys, the internet. If an agent reads an instruction hidden in a page or a document, it carries it out with all of that within reach. Fourth: the key that opens all its secrets is on the same disk as its database. Whoever can read that disk can read the secrets. Fifth: in local mode, any program on the computer that reaches its port gets in as an administrator.
And one more thing: Paperclip had serious security flaws, fixed in version 2026.416.0. If you use an earlier version, update it today. None of this means you shouldn't use it. It means that, by default, Paperclip puts no limits on your agents. You set those limits.
How do you use it with care? One: on a dedicated machine, not on the computer where you keep your life. Two: up to date and listening only on that machine. Three: for each agent, the fewest keys and a budget with a cap. Four: the sandbox on in every agent it launches. And five: check what each agent can really do.
That last one is what RosmarOps is for. Since version 0.17 it finds Paperclip on your computer and tells you which agents it can put to work without asking you, whether your version has a known flaw and whether it listens beyond your computer. Without opening its secrets and without anything leaving your computer.
To sum up: Paperclip gives you a company of agents, but by default it puts no limits on them. You are the board. A dedicated machine, the fewest keys, the sandbox, and check what they can do. See you next week.

