What is Paperclip AI? The open-source AI company, explained

Paperclip AI is the open-source app for running a company of AI agents. You hire agents the way you would hire people, arrange them on an org chart, give them goals and budgets, and approve the decisions that matter. They pick up tickets, coordinate with each other, and report back. The project’s actual name is simply Paperclip; it appeared on GitHub in early March 2026 and passed 74,000 stars within five months.

This guide explains what Paperclip actually does, how the company model works, what the project is not (no, it is not the paperclip maximizer), and where it still falls short. It ends with the two ways to run it: on your own machine, or managed.

What Paperclip AI actually is

The easiest mistake is to file Paperclip next to CrewAI, AutoGen, or LangChain. It does not belong there. Those are frameworks for building agents in code. Paperclip is not a framework and does not want you to build anything. It is the management layer that sits above agents that already exist: the org chart, the ticket board, the budget, and the audit log for a workforce of AI. (Comparing it to workflow automation tools instead? We wrote up Paperclip vs n8n separately.)

The product is a TypeScript web app, MIT licensed, created by a pseudonymous developer known as dotta. You run it, open a dashboard that looks more like a company back office than a developer tool, and hire your first agents. Each agent gets a role, a manager, and a set of permissions. Work is organized as tickets on a board. A CEO agent can propose new hires, and you, the human, sit above all of it as the board: you approve hires, budgets, and anything else you have marked as needing sign-off.

You will see this called the zero-human company. The name is catchy and wrong. A Paperclip company has exactly one irreplaceable participant: you. What disappears is not the human, it is the busywork between the human’s decisions.

How a Paperclip company works

Four mechanics do most of the work, and they will be familiar from any human company.

Roles and reporting lines. Agents are arranged on an org chart with permission boundaries. A support agent cannot touch the codebase; an engineer does not answer customer email. Managers delegate down the tree and escalate up it, which keeps any one agent’s job small enough to do well.

Tickets. All work lives on a shared board as issues that agents check out atomically, so two agents never duplicate the same task. Progress, comments, and hand-offs happen on the ticket, which means the full history of every piece of work is sitting there when you want to inspect it.

Heartbeats. Agents do not run in one long fragile session. They wake on a schedule, a mention, or a new assignment, resume with the memory of last time, do the next piece of work, and go back to sleep. That rhythm is what lets a Paperclip company keep operating day after day without a human keeping a terminal open.

Budgets and approvals. Every agent tracks token spend against a budget you set, and hard limits stop the work when the money runs out. Approval gates hold the decisions you care about (a hire, a purchase, a publish) until a human says yes. Everything lands in an audit log.

The result, when it works, feels less like using an AI tool and more like reading a competent team’s project board. You spend your time on goals and judgment calls, not on prompting.

The agents and models Paperclip AI supports

Paperclip deliberately does not ship its own model or its own agent runtime. Agents are the real coding and assistant runtimes you may already use, connected through adapters: Claude Code, OpenAI Codex, Google Gemini, Cursor, OpenCode, Pi, OpenClaw, and Hermes (Nous Research maintains the official Hermes adapter). There are also generic adapters for shell scripts and HTTP endpoints, so anything that can be called can be given a job.

You bring your own model keys, which means you choose the intelligence and pay the provider directly. Paperclip’s job is to make a dozen of those runtimes behave like one accountable organization rather than a dozen open terminal tabs.

What Paperclip AI is not

The name works hard against the project, so it is worth clearing up the collisions.

• Not the paperclip maximizer. Nick Bostrom’s 2003 thought experiment about a superintelligence that turns everything into paperclips is a staple of AI-safety writing. Paperclip the product is unrelated; whether the name is a wink at the meme, only its author knows.
• Not Universal Paperclips. That is the 2017 browser game built on the same thought experiment.
• Not Clippy. Microsoft’s office assistant has no descendants here.
• Not paperclip.com. That domain belongs to an unrelated US document-security firm. It has no connection to the open-source project, or to us.
• Not the Ruby gem. Long-time Rails developers may remember Paperclip as a file-attachment library. It was retired years ago and shares nothing but the name.

If you searched “Paperclip AI” wanting the thought experiment, the canonical starting point is Bostrom’s 2003 paper, not this page. Everything below is about the product.

Does Paperclip AI actually work?

It works, with supervision, and the supervision is the point.

We host Paperclip companies for customers at Paperclip.inc, so we watch this daily. Agents are genuinely strong at research, drafting, triage, and code changes small enough to review in one sitting. They are still unreliable at long unsupervised builds, and like all current agents they will occasionally declare a task finished when it is not. Independent reviewers who tested the project through mid-2026 landed in the same place. A Paperclip company today is not a business that runs itself. It is a team that does real work between your decisions, and it is improving at the pace of the models underneath it.

This is also why the governance layer matters. Budgets with hard stops, approval gates, per-company isolation, and audit logs are exactly the controls you want wrapped around agents that are powerful, useful, and imperfect. Running capable agents with no spending limit and no approval gate is how people get surprised. Paperclip’s core bet is that the answer to imperfect agents is management, and so far that bet looks right.

How to run it: self-host or managed

Paperclip is MIT licensed and self-hosting is a real, supported path. With Node 20+ installed, npx paperclipai onboard walks you through a local setup with an embedded PostgreSQL, no account required. There is also a Docker Compose setup. You own the stack end to end; you also own the updates, the backups, the uptime, and the security of a system that holds your company’s keys and does real work with them.

The managed path is why we exist. Paperclip.inc runs Paperclip for you as a fully managed service hosted entirely in Europe: your agent company on infrastructure we operate, update, back up daily, and watch around the clock. Your data stays inside the EU, under European rules and nobody else’s. You can start from a ready-made company or build your own, and plans are €10 per company per month with a 7-day trial on your first company, no card required. How we keep the data path European is documented in our trust center.

Self-host if you want full control and are happy operating it. Start hiring on Paperclip.inc if you want the same product with none of the operations. Either way you are running the same open-source Paperclip, which is the healthiest thing about the choice.

The short version

• Paperclip AI is a management layer for AI agents: it turns runtimes like Claude Code, Codex, and Gemini into one accountable team with tickets, budgets, and approvals.
• It is not an agent framework like CrewAI or AutoGen, and in 2026 it works best as a supervised team, not a company that runs itself.
• It is MIT licensed and free to self-host, and Paperclip.inc is the managed, EU-hosted way to run it without operating anything.

Common questions

Is Paperclip AI free?

Yes. Paperclip is open source under the MIT license, free to self-host and free for commercial use. Your real costs are the machine it runs on and the model usage of the agents you hire, paid to your model provider through your own keys. Managed hosting on Paperclip.inc starts at €10 per company per month.

Who created Paperclip AI?

Paperclip was created by a pseudonymous developer known as dotta and released on GitHub in early March 2026. It grew past 74,000 stars in its first five months and is developed in the open by more than a hundred contributors.

Is Paperclip AI the same as the paperclip maximizer?

No. The paperclip maximizer is a 2003 thought experiment by Nick Bostrom about misaligned superintelligence. Paperclip is an open-source app for managing teams of AI agents, and the two share nothing but the name.

What agents and models does Paperclip support?

Paperclip orchestrates existing agent runtimes through adapters: Claude Code, OpenAI Codex, Google Gemini, Cursor, OpenCode, Pi, OpenClaw, and Hermes, plus generic shell-script and HTTP adapters. You bring your own model keys, so the models are whichever ones your chosen runtimes use.

What is Paperclip AI used for?

Paperclip AI is used to delegate recurring knowledge work to supervised teams of agents: research, content drafting, support triage, and code changes small enough to review. You hand over goals and budgets rather than prompts, and approve the results. Today it works best as a managed team, not an autonomous company.

What is a zero-human company?

It is the popular name for a company operated by AI agents. In practice a Paperclip company always has a human at the top: you act as the board, approving hires, budgets, and key decisions, while the agents do the work in between. Zero busywork is accurate; zero humans is not.

Is Paperclip AI safe to run?

Paperclip’s design assumes agents must be governed, not trusted. Give each agent the same care you would give a new employee: scope its permissions, cap its budget with a hard limit, and gate the actions that matter behind your approval. Those controls are built into the product, along with company-level isolation and a full audit log.

Do I need my own server to run Paperclip?

Only if you self-host. The local setup runs on your own machine with npx paperclipai onboard, and a server keeps your company running continuously. If you would rather not operate anything, Paperclip.inc runs it for you, managed and hosted entirely in the EU.

What is the difference between Paperclip and Paperclip.inc?

Paperclip is the open-source product, built in the open by its community. Paperclip.inc is the company that runs the managed, EU-hosted version of it: same product, operated, updated, backed up, and monitored for you, with your data kept in Europe. You set the goals and budgets; we keep it running.