Paperclip: launch a full AI company in 30 minutes (without a single human employee)

In this article:

The problem everyone had but nobody had solved

If you use Claude Code seriously, you've been there: 20 terminals open, 20 parallel sessions, and when you come back the next morning, you have no idea who's doing what.

That's exactly the problem Dota, one of Paperclip's founders, described publicly. And it's why the tool hit 36,000 GitHub stars within just a few weeks of its early March 2026 launch.

The core idea is simple: instead of managing your AI agents directly in scattered terminals, you manage them from a centralized dashboard, like a real board of directors overseeing its teams.

You're no longer the operator doing everything. You're the board. You set objectives, metrics, expected outcomes. The agents organize themselves.

What exactly is Paperclip?

Paperclip is an AI agent orchestrator, fully open source (MIT license), available at paperclip.ing. Its tagline: "zero human companies." The idea: your company can run with AI agents that manage themselves, recruit each other, and execute continuously.

In practice, Paperclip provides:

• A web dashboard with a global view of all your agents and their current tasks

• A ticket and issue system, like a GitHub for your agents

• Heartbeats: your agents wake up on a defined schedule (every 4, 8, or 12 hours) and resume their work

• An inbox where you approve important decisions (hiring, new tasks, spending)

• A per-agent budget system to control costs

• Routines: automated recurring workflows (daily security audit, weekly reporting, etc.)

Paperclip works with Claude Code, OpenClaw, Codex, Cursor, or any compatible agent. To understand the differences between OpenClaw and Claude Code, I wrote a detailed analysis here.

The setup in 10 minutes (really)

To get started, one single command in your terminal from paperclip.ing. It launches a local server with a dashboard on localhost. By default, everything runs locally. You can migrate to a VPS to access it from anywhere.

From there, the flow is straightforward:

• 1. Create your company (name + mission)

• 2. Create your first agent (CEO, for example), pick the model: Sonnet 4.6, Opus, whatever you want

• 3. Define a first task (or keep the default template: 'recruit a first engineer')

• 4. Launch, and watch the CEO create a plan, recruit, assign tasks

What I saw in the demo: in 30 minutes, a fictional company (ProofShot, a video testimonial tool) had a CEO, an engineer, defined milestones, generated code, and a QA in the process of being recruited. Nobody touched anything in between.

The key: each agent has configuration files (soul, heartbeat, tools, agents). They know who they are, what they do, and how to operate within Paperclip.

Heartbeats: why this is the real innovation

The heartbeat concept is borrowed from OpenClaw, which had already popularized the idea of proactive agents. Paperclip integrates it natively for all agents.

At each wake-up, the agent starts fresh with updated context. It re-reads its tasks, its state, its instructions. This is what allows it to work autonomously without accumulating confusion over long sessions.

This principle of persistent memory and context between sessions is similar to what I had explored with AutoDream for Claude Code. Paperclip industrializes this approach for entire teams of agents.

Each agent can run a different model. Your CEO on Claude Sonnet, your research agent on a cheaper model, your designer on a specialized one. You optimize quality and cost per role.

Paperclip vs raw Claude Code sessions: the honest comparison

Here's what concretely changes compared to managing multiple Claude Code sessions yourself:

Skills and templates: don't start from scratch

Two things massively accelerate the setup:

First, skills.sh, an open source skills marketplace you can install into your Paperclip project by pasting a GitHub URL. Front-end design skills, web design, security audit skills... All free, verified (but still use with caution).

Then, company templates: in the official Paperclip repo, you'll find ready-to-deploy "companies." G-Stack (CEO, CTO, QA, Release Engineer, Staff Engineer), AI agency, scientific research company with dozens of specialized roles. A template comes with its agents, its skills, and its knowledge files.

If you're starting from scratch and don't know exactly which agents to create, importing a template that resembles what you want to do is far faster than building everything by hand.

My verdict

Paperclip solves a real problem: the complexity of managing multiple AI agents in parallel. It's the missing piece of infrastructure to go from "I have one agent doing one thing" to "I have an organization that runs."

What I find particularly convincing: the board vs operator philosophy. You no longer micro-manage agents. You set high-level objectives, approve important decisions, and let the organization execute. It's a genuine shift in how we interact with AI tools.

The logical next step after Paperclip is integrating it into a real content or marketing strategy. I had explored how to build a complete AI marketing team with Claude Code, and the two approaches complement each other well.

What deserves attention: Paperclip is still young (a few weeks old). Features like routines are in beta, the secrets/environment variables system is poorly documented. You need to use Claude Code itself to navigate the gray areas.

My rule: if you regularly manage more than 2 to 3 Claude Code agents in parallel, Paperclip is worth the setup time. Below that, the terminal is enough.

36,000 GitHub stars in a few weeks is no accident. It's validation that the market was waiting for exactly this.