30,000 GitHub stars in under three weeks.
That's how fast Paperclip blew up. It's an open source tool for running Paperclip AI agents like actual employees: org charts, roles, budgets, hiring plans, performance tracking. The tagline is "zero-human companies."
Greg Isenberg sat down with Dotta, the pseudonymous creator, for one of his first public walkthroughs. They spun up a whole AI company live on the call: a CEO, a founding engineer, a QA agent, and a video editor, all working on a finance app Greg picked off his idea browser.
I watched the whole 46 minutes so you don't have to. Here's what Paperclip actually is, plus the one mental model from the episode worth stealing even if you never install it.
What Paperclip Actually Is
Think of the spectrum of AI tools right now.
On one end you've got fully automatic tools: hand over your credit card, they generate business ideas and run with them. On the other end you've got coding agents in 15 open tabs while you babysit pull requests.
Paperclip sits in the middle. You define business goals, hire a team of agents, approve their plans, and they work. Every task runs through an issue tracker, and one agent owns one task at a time, so they stop stepping on each other.
The part that made it spread: it's bring-your-own-bot. Claude Code, Codex, Cursor, OpenClaw, or OpenCode, which opens up any model on OpenRouter.
Dotta's advice on models is simple:
→ Frontier model for your CEO agent. That's where judgment lives.
→ Cheaper models for grunt work.
→ Check OpenRouter for free models. Some are free for weeks at a time, and some tasks don't need more.
It runs on your local machine today. Cloud hosting is coming, but local is how almost everyone uses it right now.
The Origin Story: 30 Claude Code Windows and a Busted Budget
Dotta built Paperclip out of a mess every agent-heavy founder will recognize.
He'd have 20 or 30 Claude Code windows open at once, building his companies. He'd set them running over the weekend and come back Monday with no idea what any of them did. Worse: they'd spent his token budget doing it.
(Been there. It's a special kind of dread, opening tab 23 to find out what it cost you.)
So Paperclip tracks everything. Monthly token spend per company. Every task an agent touches, logged as an issue you can open, read, and comment on. In the live demo the spend counter sat at $0 the whole time because the agents ran on Dotta's existing Claude and Codex subscriptions instead of API credits.
That accountability layer is the whole pitch. The agents aren't smarter, you can just finally see what they're doing.
The Live Demo: Hiring a Company for a Finance App
Greg picked an idea off his idea browser: a finance app that builds money habits in 3 minutes a day. They named it Moola. Then they hired it a staff, live.
The flow looked like this:
- Paste the company goal into Paperclip.
- Paperclip spins up a CEO agent, which writes a hiring plan and breaks the roadmap into tasks.
- The CEO asks permission to hire a founding engineer. You approve it from an inbox. (Once you trust the system, you can let agents hire agents automatically.)
- Five live tasks appear: scaffolding, CI, auth and onboarding, the core daily loop, progress tracking.
- They hired a QA agent to review the engineer's work, then a video editor with the Remotion skill attached, to plan a 60-second animated pitch video for Twitter.
The mindset shift is the point. You're not writing prompts for jobs to be done. You're the board. You tell the CEO "here's the MRR target, make a plan," and you review what comes back.
Dotta's power-user tip: whenever you don't know how to do something in Paperclip, just ask your CEO agent to do it. It knows the tool better than you do.
The Memento Man Rule (Steal This One)
Best part of the episode, and it applies to any agent setup you run, Paperclip or not.
Dotta's mental model: your AI agents are the guy from Memento. They wake up extremely capable. They can code, they can browse, they can spend your money. But they have no idea who they are, where they are, or what they were doing.
Your job is to leave them the tattoos and Polaroids.
In Paperclip that's a "heartbeat" checklist every agent runs on wake-up: confirm who you are, read today's plan, find your assignments, break them into tasks, write your memory down. Memory defaults to plain files using the PARA method, with skills borrowed from Nat Eliason's setup.
And when an agent does something you hate? Don't rage-quit. Open its persona and add one line. Dotta's example: his engineer kept shipping work with no definition of done, so he added "set a success condition for every task and ask QA to review it." Save it. The output gets better.
He was honest about the ceiling too: "AI can do everything except know your values." The models bring the skills. You have to write down the taste yourself, same as you always had to with human hires.
Skills Are the New Hiring Market
Agents in Paperclip get hired with skills attached, and there's already an economy forming around them.
The main marketplace mentioned is skills.sh. The Remotion skill (for programmatic video) has over 100,000 stars. So "hire a video editor" becomes: create an agent, paste the Remotion skill URL, done.
Greg asked the uncomfortable question: how do you know a skill isn't malicious? Dotta didn't dodge it. Nobody has fully solved that. skills.sh runs security audits and shows badges, and star counts are a signal, not a guarantee. Treat it like installing npm packages in 2015. Be careful.
The wilder feature is importable companies, which was days from shipping when they recorded. You install someone else's entire org into your Paperclip. Gary Tan's "G stack." A 60,000-star repo with over 100 prebuilt agents. A full game studio, which Dotta imported and asked for a bullet hell game in Godot, Vampire Survivors style. It started planning, and he paused it because it was eating too many tokens.
Proven org charts you can clone like a repo. That's a genuinely new thing.
Who's Actually Using It (And the Honest Caveat)
Three weeks in, the real usage isn't zero-human companies. It's normal businesses bolting agents onto existing operations:
→ A security review company used Paperclip to run security reviews on Paperclip itself, and now manages client audits with it.
→ A dentist organizes his foundation work and family admin through it.
→ A roofing company is exploring agents that pull hail data and satellite imagery to hand sales reps better leads.
Now the caveat, and credit to Dotta for volunteering it: Paperclip has made $0 in revenue. The 300-agent mega-orgs you can import? "Completely unproven at this point," his words. Nobody has run evals on them. The companies making money are the boring ones using three agents, not thirty.
The team behind it is real, though. Dotta (previously of Forgotten Runes in the NFT era) plus two co-founders: Devin Foley, early at Slack and Figma, and Scott Tong, former head of product design at Pinterest. And a community sending more pull requests than they can merge, 500 open at recording time.
What's Coming: Routines and Maximizer Mode
Two roadmap items stood out.
Routines are repeating issues with full tracing. In the demo they built one live: every day at 10am, read the last 24 hours of merges and draft a community changelog for Discord, celebrating contributors by name. Same task, every day, with a log of what it did and what it cost. If you've ever duct-taped that together with cron and prayer, you get it.
Maximizer mode is the ambitious one. You stop caring about token spend and tell the CEO: someone works on this until it's done. Build the game until it's playable. (Yes, "paperclip maximizer." The name is a bit on the nose.)
Greg asked the question hanging over every agent tool: does Paperclip still exist in a year, or do the models eat it? Dotta's bet is that even when agents replace headcount, you still need the layer that manages taste and organization at scale. Maybe. It's the same bet every orchestration startup is making.
My Take for Bootstrapped Founders
You don't need a 30-agent company. You need one agent doing one job you can verify.
I run AI agents across my own stuff, and the pattern that works matches what I found when I broke down how Allie K. Miller runs her company with 34 AI agents: what matters is writing down what good output looks like before you delegate, not how many agents you run.
So if Paperclip tempts you, start like this:
- One company, one goal, stated in MRR or a shipped feature.
- One CEO agent on a frontier model, one worker on something cheap.
- One routine you'd otherwise do manually every day.
- Add a rule to the persona every time output disappoints you. Six months of those rules is something nobody can copy in a weekend.
And before you install a single skill, read up on how Claude skills actually work. Ten minutes there saves you a weekend of debugging an agent that "forgot" its job.
FAQ
What is Paperclip AI?
Paperclip is an open source agent orchestrator for running teams of AI agents like a company, with roles, hiring plans, issue tracking, and token budgets. It hit 30,000 GitHub stars in its first three weeks. You bring your own agents (Claude Code, Codex, Cursor, OpenClaw, or OpenCode) and it coordinates them so one agent owns one task at a time.
Is Paperclip free?
The software is open source and free to self-host on your local machine. Your real cost is inference: agents run on your existing Claude or Codex subscriptions, or on API credits through OpenRouter. Paperclip tracks that spend per company, which is half the reason it exists. A hosted cloud version is planned.
Which AI models work with Paperclip?
Any of them, effectively. Native support covers Claude Code, Codex, Cursor, and OpenClaw, and OpenCode support means you can run any model listed on OpenRouter. Dotta recommends a frontier model for your CEO agent and cheaper or free models for routine tasks.
Can Paperclip really run a zero-human company?
Not yet, and the creator says so himself. Paperclip has made $0 of its own revenue, and the big importable agent orgs are unproven. The people getting value today use it to manage a handful of agents inside an existing business: security audits, lead research, content routines. Treat "zero-human" as the direction, not the current state.
Want more teardowns like this, straight from founders doing the numbers out loud?
I interview bootstrapped SaaS founders making $100K to $10M a year about exactly how they run their companies, agents included.
Listen to the Profitable Founder Podcast →