← Back to Blog

Paperclip AI Review: Setup, Features + How It Works

by Tom
SEO

Paperclip AI is a free, open-source orchestration platform that lets you build autonomous AI teams, where you hire AI employees, give them job titles, assign them tasks, and they run operations on their own. Think of it as the company that wraps around tools like Claude Code or Codex. Those tools are the individual workers. Paperclip is the org chart, the CEO, the management layer, and the Kanban board that ties them together.

I'm Tom. I've been in the automation space for over eight years, and I tested Paperclip over a few days on a VPS to figure out whether it's a real productivity tool or another hype cycle. Below is a hands-on Paperclip tutorial, my honest Paperclip review, the standout features, where it falls short, and how I'd actually set it up if I were starting today.

Watch the full Paperclip review and tutorial

Video walkthrough: Watch on YouTube: Paperclip Is the Craziest AI Tool I've Ever Seen (Full Review + Tutorial). Everything below is the written companion to that walkthrough. Read it for the deep dive, watch the video to see Paperclip in action.

What is Paperclip?

Paperclip is an open-source AI orchestration platform that lets you build what its creators call "zero human companies." Instead of running a single AI agent, you spin up a whole organisation: a CEO at the top, C-suite executives below, and engineers and staff under those. Each agent gets a job title, a system prompt, instructions, skills, and a set of tasks. They communicate, delegate, and report back to you.

Free, open-source, and over 44,000 developers on GitHub. The simplest way to think about it: if Claude Code or Codex are individual AI employees, Paperclip is the company those employees work inside.

How Paperclip works

Paperclip runs as a self-hosted application. You can install it on your own machine, a Mac Mini, or a VPS. Once it's running, you log into a web interface that looks like a project management tool crossed with an org chart. From there, the structure works top-down.

The hierarchy: company, project, agent, task

Everything is structured top-down. You set a company mission. Inside the company you create projects. Each project has a goal. You then hire agents, starting with a CEO, and assign them issues, which are essentially tasks. The CEO orchestrates the other agents and delegates work down the chain.

Heartbeats and autonomy

Only the CEO has a heartbeat, which is a recurring trigger that runs on whatever interval you set. The default is 3,600 seconds, which is one hour. When the heartbeat fires, the CEO checks for new issues and decides what to do with them. You (or other agents) create the issues, the CEO assigns them, and the system runs without you needing to manually trigger every step. That's where the "autonomous" claim actually starts to hold up.

Adapters: the LLM behind each agent

Each agent runs on top of an adapter: Claude Code, Codex, or any LLM you want to plug in. Under the hood, every agent is essentially an instance of Claude Code (or whichever model you picked) with its own system prompt, AGENTS.md file, skills, and configuration. So if you have five agents plus a CEO, you've got six concurrent instances of Claude Code running.

What I tested

I deployed Paperclip on a Hostinger VPS using their one-click install (KVM2 plan, $8.99/month). Deploy took a couple of minutes. Then I ran the full setup: connected Claude via SSH, named the company "AI Growth Lab," wrote the mission and goal, hired a CEO, and told it to hire a founding engineer.

From there I built a real workflow: an autonomous social media content pipeline. I created four specialist agents: a content researcher, a hook writer, a post writer, and a quality checker. Each one got the system prompts, MD files, and copywriting principles I already use manually inside Claude Code (humanizer skill, brand voice references, copywriting checklists). I gave the CEO an issue: pull research from Reddit and YouTube, narrow to three winning ideas, generate hooks, write the posts, run every draft through the humanizer pass.

Then I ran the heartbeat. The CEO picked up the issue, delegated to each agent in order, and the agents talked to each other through the system. The output was a LinkedIn post with the headline "20 best employers probably hiding their AI work for...", which is actually a usable hook. That tracks, because the underlying skill files are the same ones I use to write content that pulls millions of views. The orchestration worked. I then saved the whole thing as a daily routine, so it now runs in the background automatically.

Standout Paperclip features

Five things stood out from the test:

  1. The org chart is the product. You're not building one agent, you're building a company. CEO at the top, C-suite below, engineers underneath. The org chart isn't decorative. It's how delegation actually flows.
  2. Heartbeats turn agents into background workers. The CEO can run on a schedule, check for new issues, and dispatch work without you. It's the closest thing I've seen to a true "set it and forget it" agent system.
  3. Routines for daily workflows. You can save any process (like my content pipeline) as a routine, set it on a daily trigger, and Paperclip will run the whole multi-agent workflow on schedule. Mine now ships content options into my inbox every day.
  4. Skills and MCPs work across agents. You can install Claude Code skills inside the VPS terminal, paste them into Paperclip, or pull them from a skills marketplace. Once installed, agents can use them. Same for MCPs.
  5. Import and export companies. You can export a working organisation and someone else can import it. That means the most useful thing about Paperclip might end up being a marketplace where people sell pre-built AI companies that already work for a specific use case.

Where Paperclip falls short

I'll give you the criticism straight, because anyone telling you this is plug-and-play hasn't actually used it.

First, the setup is not for beginners. The marketing makes it sound like "define the goal, hire the team, approve, and run." In reality, you need to be comfortable with a VPS, SSH, Docker containers, environment files, and at least a basic understanding of how Claude Code skills and MCPs work. If you've never opened a terminal, you'll struggle.

Second, the outputs are only as good as the context you put in. Paperclip orchestrates. It doesn't write your content or do your research. Every system prompt, every AGENTS.md file, every skill has to be built and tuned by you. Garbage in, garbage out, but with extra steps.

Third, security is your problem. The platform doesn't wrap your VPS in best practices. You'll want something like Tailscale to lock down access so only you can connect. API keys belong in the environment file, never pasted into the chat or an issue comment.

Fourth, costs scale fast. Each agent is its own instance of Claude Code. Five agents plus a CEO is six concurrent instances burning tokens. If you're using a Claude subscription, you'll hit your limits quickly with autonomous workflows running in the background.

Fifth, and this is important: Anthropic's updated rules say you cannot connect your Anthropic subscription to third-party harnesses like Paperclip or Open Claude. The video shows me doing it to demonstrate the SSH process, but I don't recommend it. Use an Anthropic API key instead.

Paperclip vs other AI tools

Paperclip vs Claude (Claude Code)

Claude Code is one AI agent doing one job at a time, in your terminal. Paperclip is a layer above Claude Code. It spawns multiple Claude Code agents, gives them roles, lets them communicate, and runs them on schedules. You don't pick one or the other. Paperclip uses Claude Code as its workforce.

Paperclip vs ChatGPT

ChatGPT is a chat interface for prompting one model. Paperclip is an orchestration platform for running an autonomous AI team in the background, with delegation, heartbeats, routines, and a Kanban board. They're not the same category. If you want a conversation, use ChatGPT. If you want a system that runs while you sleep, use Paperclip.

Paperclip vs Open Claude

Open Claude is another harness for running Claude, closer to Claude Code itself. Same logic as the Claude Code comparison: Paperclip sits above harnesses like Open Claude as the orchestration layer.

My take: Paperclip is a step up from Claude Code and Open Claude. It's the next layer of the AI automation stack.

How to set up Paperclip

Here's the actual setup process I used. Don't worry if you've never used a terminal. Follow each phase in order.

Phase 1: Choose where to run it

You have three options: your own laptop, a Mac Mini, or a VPS. I don't recommend running it on your daily-driver laptop because it has access to all your files. A Mac Mini works as a sandboxed environment but you need to keep it running 24/7. A VPS is the cheapest and cleanest way to get started. I used a Hostinger KVM2 at $8.99 a month, with a one-click Paperclip install in their Docker catalog.

Phase 2: Deploy Paperclip

Inside Hostinger (or any VPS provider with Docker support), go to Docker Manager, then Catalog, search for Paperclip, set your admin email and password, and hit deploy. Takes a couple of minutes. If you're using your own VPS, install via Docker manually.

Phase 3: Add your API keys to the environment file

Do not paste API keys into the Paperclip chat or comments. That exposes them. Find the .env file in your Docker manager, add ANTHROPIC_API_KEY and any other keys you want Paperclip to use, and save. This is also how you give it access to additional services.

Phase 4: SSH into your VPS and install Claude Code

From your terminal, SSH into your VPS as root using the password you set in the hosting dashboard. Run docker ps to find the Paperclip container ID, then docker exec -it [CONTAINER_ID] /bin/bash to drop into the container. From inside, run claude to install and authenticate Claude Code. This is what makes skills and MCPs available across all your Paperclip agents.

Phase 5: Lock down security

Before you put anything serious in there, secure the VPS. Ask Claude how to set this up. Most people end up using Tailscale so the VPS is only reachable when connected to their Tailscale network. Don't skip this.

Phase 6: Set up the company, mission, and CEO

Open Paperclip in your browser, log in, and name your company. Write a clear, specific, concise mission and goal. I asked Claude Code to draft mine using my full company context. Then create your first agent: the CEO. Pick the adapter type (Claude Code), test it, and confirm it goes green. Set the CEO's heartbeat interval (the default of one hour is fine to start).

Phase 7: Hire your team and start small

Don't try to build a 10-agent organisation on day one. Start with the CEO and one engineer. Give the CEO a small first task. "Hire a founding engineer" is a good starter. Approve the hire from your inbox. Then add a project, hire the specialists you need, and give each one a tuned system prompt and AGENTS.md file. Iterate from there.

Verdict

Paperclip is the most exciting orchestration tool I've seen in a while. The ability to run a full AI team that delegates work, checks each other's outputs, and runs operations around the clock without you lifting a finger is genuinely a step up from working with single agents. I'm only scratching the surface here, and I'm bullish on where this goes, especially the import/export angle, which could turn into a real marketplace for pre-built AI companies.

Who it's for: people already comfortable building with Claude Code, who have specific workflows they run manually and want to automate end-to-end. If you've already got skills, prompts, and processes dialled in, Paperclip is the missing orchestration layer. Who it's not for: total beginners who haven't touched a terminal, or anyone expecting a no-code drag-and-drop builder. The setup will frustrate you and the outputs will be weak because the inputs will be weak. Get comfortable with Claude Code first, then come back to Paperclip.

Paperclip AI FAQ

What is Paperclip?

Paperclip is a free, open-source AI orchestration platform that lets you build autonomous AI teams. You set a company mission, hire AI agents (CEO, engineers, specialists), assign tasks, and the system runs delegation and execution on its own.

Is Paperclip free?

Yes. Paperclip itself is free and open-source, with over 44,000 developers on GitHub. You'll pay for the infrastructure you run it on (a VPS like Hostinger starts around $6.50 to $8.99 a month) and for whatever LLM you plug in via API.

How do I install Paperclip?

The fastest path is a VPS with a Docker one-click install. I used Hostinger's Paperclip catalog item. Alternatively, you can self-host on a Mac Mini or your own machine via Docker. After deploy, you log in with the email and password you set, add your API keys to the environment file, SSH into the container to install Claude Code, and start building your company.

Is Paperclip safe to run?

Running an autonomous AI organisation on your daily-driver laptop is not the best idea, because it has access to your files. Use a VPS or a sandboxed Mac Mini, lock it down with something like Tailscale so only you can connect, and keep API keys in the environment file rather than pasted into the chat.

Is Paperclip better than ChatGPT?

They're not the same category. ChatGPT is a chat interface. Paperclip is an autonomous orchestration platform that runs multiple AI agents on a schedule. If you want a conversation, use ChatGPT. If you want a system that delegates and ships work in the background, use Paperclip.

Does Paperclip work on Mac and Windows?

Yes. The setup is essentially the same on Mac, Linux, and Windows (PowerShell). The only difference is which terminal you use to SSH into your VPS or run Docker locally.

Can I connect my Claude subscription to Paperclip?

No, and this is important. Anthropic's updated rules state you cannot connect your Anthropic subscription to third-party harnesses like Paperclip or Open Claude. Use an Anthropic API key instead. You'll be able to track costs in Paperclip directly.

Ready to build production-ready AI systems?

If you want to go deeper than "I installed a tool" and actually build production-ready AI systems that generate results for your business, the 30-Day Claude Code Challenge takes you through an interactive Claude Code course inside Claude Code. You'll build your first web app, your first automation, and walk through the full process of translating an idea into a working system. If the cohort is open, applications are live now.

Not ready for the challenge yet? Grab the free Claude Code Blueprint. It's the starting point most people use before they commit to building.