Paperclip: the app that wants to run a company on AI agents
Every so often a repo shows up whose README makes you do a small double-take. Paperclip (the paperclipai/paperclip one, not the ancient Rails file-attachment gem) is that repo for me this month. Its pitch: it's "the app people use to manage AI agents for work" — a Node.js server and React UI that orchestrates a whole team of agents to, and I quote, run a business.
So I went digging. Here are the field notes.
What it actually is
Paperclip is open-source orchestration for teams of AI agents, built by Paperclip Labs, Inc. and released in early 2026 by a pseudonymous developer. It reportedly racked up tens of thousands of GitHub stars within weeks, which tells you more about the current appetite for agent hype than about whether it works — but the code is real and MIT-licensed.
The mental model it sells is organizational, not tool-shaped. The README's line is the clearest summary I found:
If OpenClaw is an employee, Paperclip is the company.
So instead of babysitting one agent in a terminal, you get org charts, roles, reporting lines, goals that tasks trace back to, budgets, governance, and an audit trail. It's the management layer, not the agent itself.
The architecture, briefly
It's TypeScript top to bottom (~97%), React on the front, embedded PostgreSQL for local dev (or point it at your own), and it's happy running fully self-hosted with no Paperclip account required. A few pieces stood out as genuinely thoughtful rather than decorative:
- Bring your own agent — Claude Code, Codex, Cursor, plain bash, or an HTTP bot. It doesn't ship its own model.
- Heartbeat execution — agents get scheduled wake-ups with atomic checkout of work, so two agents don't grab the same task and you don't get runaway double-work.
- Cost control — per-agent monthly budgets with warning thresholds and hard stops that pause agents and cancel queued work on overspend. This is the feature I'd actually want.
- Multi-company isolation, routines (cron/webhook/API triggers), encrypted secrets, and a ticketing system with immutable logging.
Getting it running
Onboarding is a single command. There's a public-registry escape hatch too, because private npm registries apparently trip it up:
# fastest path
npx paperclipai onboard --yes
# if a private registry blocks install
npx --registry https://registry.npmjs.org paperclipai onboard --yes
# manual / dev setup
git clone https://github.com/paperclipai/paperclip.git
cd paperclip
pnpm install
pnpm dev # API + UI, watch mode, at http://localhost:3100
You'll want Node 20+ and pnpm 9.15+. One thing to flag: telemetry is on by default. It claims not to collect prompts, secrets, or personal data, and you can switch it off with PAPERCLIP_TELEMETRY_DISABLED=1 or the usual DO_NOT_TRACK=1. Worth doing on principle before you forget.
Where the pitch cracks
Here's the honest part. The "zero-human company that just runs" framing is aspirational, and the reviews that actually built something with it say so plainly. The recurring complaints: there's no quality filter in the loop, so nobody catches that a stat is hallucinated or that the shiny generated website doesn't render on mobile. Brainstorming and email drafts land fine; the autonomous output is where it wobbles.
And it is not no-code. You need to be comfortable with the command line, Node, pnpm, env vars, and reading docs. The marketing understates how many real hours you'll spend on prompts and governance before anything approaches "runs itself."
My read: Paperclip is a genuinely interesting piece of infrastructure — the budgets, atomic work checkout, and audit trail are the sort of unglamorous plumbing that agent swarms actually need. I just wouldn't hand it a company card and walk away. Treat it as a coordination substrate you supervise, not a replacement for judgment, and it looks a lot more useful than the tagline suggests.