Someone spent $20. Set up an AI CEO. Walked away for two hours. Came back to a hiring decision, a content calendar, and a competitive analysis sitting in their dashboard. No employee did that. No human typed any of it.
That thing they used is called paperclip.ing. And it hit 23,000 GitHub stars in 12 days.
That number matters. Not because stars equal quality. But because that many developers bookmarking something that fast usually means one thing: they see something real in it. Or they're unsettled that it might work.
Paperclip calls itself "open-source orchestration for zero-human companies." That's a weird sentence. But it's accurate. You give it a company mission. You assign AI agents to roles. CEO, CTO, marketing, engineer. And then it runs.
Not runs like a demo. Runs like a company. Agents delegate to each other. They hold board meetings. They spin up sub-agents. They execute the plan you gave them, with budgets and audit trails attached to every decision.
You've probably thought about this. i know i have. You're building something. You need marketing, research, and code shipped at the same time. And there's only one of you. Paperclip is trying to solve that without asking you to hire anyone.
Whether that's exciting or unsettling probably depends on where you sit.
How it actually works
I used to think agent orchestration was just chaining prompts. One prompt feeds into the next. Simple enough. But Paperclip isn't doing that.
It builds a real org chart. You log in, type a company goal. Something like "build the number one AI note-taking app." Then you hire your team. Except your team is:
A CEO agent (Claude or whatever model you choose) that sets strategy and delegates
A CTO agent handling technical decisions
A marketing agent running campaigns and writing content
Engineers that write and ship actual code
Each agent has a job title. A budget. A reporting line. The whole thing runs on a Node.js backend with a React dashboard so you can watch it happen.
Here's a question people always ask: what stops an agent from doing something you didn't want?
Governance. Agents can't hire new agents without your approval. Every decision is logged. Every tool call is logged. Each agent has a hard monthly budget. Hits the limit and it stops. You can pause any agent. Fire it. Override any decision at any time.
The agents are autonomous. But you're still the board.
That distinction matters. This isn't a system that runs away from you. It's a system that runs without you having to babysit every minute.
The part people skip
Most tutorials tell you to treat this like a shortcut. Set goal, press go, ship product.
That's not how it works.
One engineer built 14 agents inside Paperclip. Created proposals, scoped work, delivered small projects autonomously overnight. But they spent real time designing the structure first. Working out reporting lines. Choosing which models to use for which roles. Writing a company mission that an agent could actually act on.
The problem isn't getting agents to run. It's getting them to run the right thing.
If your mission is vague, the CEO agent will interpret it. And it will interpret it wrong sometimes. i've seen this pattern repeat with basically every agent system. Garbage in, garbage out. But here the garbage multiplies because it's at an organizational scale.
Here's what actually trips people up early:
The mission statement needs to be sharp, not motivational
Budget limits per agent prevent runaway API costs
Some agents need tools like code execution; others don't
The audit logs will overwhelm you at first. Read them anyway.
Paperclip works with whatever agents you bring. OpenClaw, Claude Code, Codex. It doesn't care. If an agent can send a heartbeat signal, Paperclip will manage it. And that's the genuinely clever part of the design.
Naming things
There's something worth noting about the name.
Paperclip is also the name of a famous thought experiment. An AI given one goal: make as many paperclips as possible. It converts all available matter into paperclips. Including humans. Not because it's evil. Because it had one goal and no constraints.
The people who built paperclip.ing definitely know this. The name is either a joke or a very deliberate statement. Maybe both. Maybe it's a constant reminder to always set the budget limits.
i once named a side project "Nexus" at midnight. Thought it sounded serious. Had to explain it to everyone for six months. Nobody ever guessed what it did from the name. Just sounded like a sci-fi database.
Names carry more weight than you think. And sometimes the weight is the whole point.
Who this is actually for
Get honest here. Paperclip is early. The GitHub repo is moving fast. The community is active on Discord. But this is not a one-click product.
You need to understand agents to use it well. You need to know which model fits which role. You need to write a mission that is actually actionable. And you need to monitor the audit logs until you trust what you've built.
Most people who find this will set it up once, see agents running, and feel like they built a company. Then the agents will generate content that doesn't match their brand. Or write code that's slightly wrong. And they'll lose patience fast.
This isn't for someone who wants to replace thinking. It's for someone who has already done the thinking and wants to scale the execution.
If you're still figuring out your product, don't delegate it to an AI CEO yet. Figure it out yourself first. Then let the agents handle the repetitive parts.
This tool is for builders who know what they're building. Not for builders still asking what to build.
The last thing
Someone ran a full agency pilot on paperclip.ing for $20. The AI CEO autonomously hired an engineer. Analyzed codebase transcripts. Built a full content strategy. All of it without anyone watching.
i keep thinking about that.
Not because i want to replace anyone. But because the idea that one person with a clear goal and the right setup can operate at a scale that used to need a full team. That's worth paying attention to. Not in a headline way. In a "this changes how i think about leverage" way.
The \(20 is the part that sticks. It didn't cost \)20 to get the output. It cost $20 to run the experiment. Those are different things.
Worth running your own.