Paperclip is an open-source orchestration app for teams of AI agents: a Node.js server with a React dashboard that hands each agent a goal, a budget and a place in a reporting structure. It is aimed at developers and technical teams who have gone past a single coding assistant and now keep several agents running at once, and who want one screen showing what those agents are doing and what they are costing. The project is TypeScript, MIT-licensed and self-hosted, and it has drawn about 86,000 stars and 15,000 forks since it first appeared in March 2026.
What it does
The README puts its own pitch in one line: if OpenClaw is an employee, Paperclip is the company. Instead of treating an agent as a session you babysit in a terminal tab, Paperclip treats it as a hire. You add agents — the materials describe roles such as a CEO, engineers and marketers — from any provider you already use, and each one gets three things attached to it: a role, a budget and a boss.
From there the app behaves roughly like a task manager, except the people in it are agents:
- Work is assigned as goals and tasks, and progress against them is tracked rather than inferred from scrollback.
- Spend is tracked per agent, and a budget is an actual limit: agents stop when they hit it.
- Decisions are logged so you can go back and review what an agent chose and why.
- The org chart itself is a first-class object, so agents report to other agents rather than all reporting to you.
The framing the project keeps returning to is that you manage business goals, not pull requests.
How it works
Paperclip runs as a Node.js server with a React user interface in front of it. The server is the coordination layer: it holds the org chart, the budgets, the goal assignments and the governance rules, and it drives the agents accordingly. You bring your own agents rather than getting new ones from Paperclip — it is the management layer, not the worker and not the model.
The result is deliberately unremarkable to look at. The README says it looks like a task manager, and the video shows exactly that: a board where tasks get assigned, work gets ticked off, costs accumulate against a cap, and a log fills up underneath. The interesting part is that the underlying objects are org charts, budgets, goal alignment and agent coordination rather than tickets and sprints.
Getting started
Because Paperclip is a self-hosted server plus a web UI, running it means standing it up yourself and then connecting the agent providers you already have credentials for. The README opens with a quickstart section, and the project keeps separate documentation at docs.paperclip.ing alongside a website and a Discord for questions. The licence is MIT, so both self-hosting and modification are unrestricted, and the repository was last pushed in September 2026.
Two practical notes before you start. First, Paperclip is useful in proportion to how many agents you have: it is a coordination tool, and one agent does not need coordinating. Second, budgets only mean something if the agents you attach report their costs, so plan on wiring up the providers you actually spend money with rather than a toy setup.
When to use it / when not
Reach for it when the problem you have is the one the video opens with: twenty agent tabs open and no idea which is working on what, what has already been decided, or how much the afternoon cost. Paperclip is also a reasonable fit when you need an audit trail — the decision log exists precisely so a person can review agent behaviour after the fact — or when uncapped spend is the thing keeping you from letting agents run unattended.
Skip it if you run one assistant in one terminal and are happy there; an org chart, a budget system and a governance layer are pure overhead at that size. Skip it too if what you want is the agent itself. Paperclip ships no model and no worker, so on its own it does nothing until you supply agents for it to manage.
Alternatives
The comparison the project draws itself is with OpenClaw, and it draws it as a complement rather than a replacement: the individual agent versus the organisation the agents sit in. In practice the real alternative most people are using today is the one the video mocks — a row of terminal tabs, plus a human keeping score in their head or in a general-purpose task tracker. That works until the agent count grows, which is the point at which this project starts to make sense.
Paperclip is worth taking seriously if you have already crossed from using an AI agent to running several of them, and have noticed that the hard part is no longer capability but supervision: who is doing what, under whose authority, against which goal, and at what cost. It makes no claims about making agents smarter. It claims that agents need management, and it provides the management — open-source, self-hosted, and clearly opinionated about the company metaphor. If that metaphor matches how you already think about your agents, it will feel obvious; if it does not, this is a lot of structure to adopt.