Here's a problem that didn't exist two years ago: you deploy a team of AI agents to run parts of your business, and then you have no idea what they're doing.
One agent books a meeting slot that conflicts with another agent's calendar hold. Two agents duplicate a piece of work because neither knew the other had checked it out. A third agent makes 47 API calls in a loop chasing a goal that shifted three days ago. Your bill arrives and it's twice what you expected.
This is the multi-agent chaos problem. Nobody solved it until Paperclip.
What Paperclip is
Paperclip is an open-source Node.js orchestration framework for running teams of AI agents. It launched in early 2026 and hit 14,200 GitHub stars and 1,600 forks in its first week — the kind of traction that happens when developers recognise a solution to a problem they've been hacking around for months.
MIT-licensed, fully self-hosted, no Paperclip account required. A single Node.js process with embedded Postgres runs locally. For production, you point it at your own Postgres and deploy however you like. One deployment can run unlimited companies with complete data isolation.
The core idea: treat your agents like employees, and build the infrastructure to manage them accordingly.
Org charts for your agents
Paperclip gives your agents job titles, reporting lines, and described responsibilities. They're not just a pool of processes that run tasks — they're a team with structure.
An agent knows its role, who it reports to, what it's responsible for, and what the broader company goal is. Context flows up from the task level through project goals to company objectives. An agent never has to guess whether what it's doing matters — it can see the whole chain.
This sounds like overhead. In practice, it's what makes agents actually useful at scale. An agent that understands the "why" behind a task makes better decisions than one that's just executing a prompt.
Hard budget limits, not "be careful"
Every agent has a monthly budget. When it's spent, the agent stops. Not "slows down" or "sends you a notification" — stops.
Atomic task checkout prevents double-work: when an agent picks up a task, it's locked. Another agent can't grab the same task and start duplicating effort. This is the kind of race condition that burns money silently in naive multi-agent setups.
If you've ever watched a poorly-configured agent loop on a problem and wondered how much it was going to cost you before you noticed — this is the feature that fixes that anxiety.
Governance that doesn't get in the way
Every config change is versioned. Bad changes can be rolled back. Approval gates are enforced — if something requires human sign-off before proceeding, the system enforces that rather than trusting the agent to know when to pause.
The audit trail is automatic. You can see exactly what each agent did, when it did it, and why (based on the context it was working from). When something goes wrong, you don't have to reconstruct what happened from logs — it's all there.
Context that survives session boundaries
This one is underrated. Most agent frameworks start fresh every time. An agent resumes a task by re-reading all available context from scratch, which is slow, expensive in tokens, and prone to drift if context has changed.
Paperclip persists agent context across sessions. An agent picks up exactly where it left off — same mental model, same understanding of where the task sits in the broader project. No cold-start amnesia.
Getting started
npx create-paperclip-app my-company
cd my-company
npm start
That's it for local use. Embedded Postgres, no external dependencies. The dashboard runs on localhost and gives you a visual org chart of your agent team, task queue, budget status, and audit log.
For production, set DATABASE_URL to your Postgres instance and deploy the Node.js process wherever you run things. It scales to multiple companies on a single instance with full data isolation between them.
The GitHub repo is at paperclipai/paperclip. Documentation covers agent definition, goal setup, budget configuration, and the approval gate system in detail.
Who this is for
If you're running more than one AI agent for any business purpose, Paperclip is worth 20 minutes of your time. The budget controls alone justify the setup if you've ever had an unexpected API bill.
If you're building out a proper AI-augmented business — where agents handle specific domains and you want them coordinated, accountable, and cost-controlled — Paperclip is the operating system that was missing.
The "zero-human company" framing in the marketing is a bit much. But the infrastructure underneath it is genuinely solid. You don't have to automate your entire business to benefit from an org chart and a budget.
Start with one agent, one goal, one monthly budget. See what breaks. Paperclip makes it visible.