Yes, the framework is open–source. Your costs are the VPS ($7–$30/month depending on plan) and either model API tokens or your existing subscriptions.

Can I use OpenClaw instead of Claude Code?

Yes — Paperclip ships with an OpenClaw Gateway adapter alongside Claude Code, Codex, Cursor, Pi, Gemini, and OpenCode. Config is a little more involved, but the model is the same.

How many agents is too many?

Practically, 10–15 is the upper end you can supervise. The platform doesn't cap depth or breadth, but bigger org charts become wallpaper.

Can agents really hire each other?

Yes. The CEO can request a new hire; you approve from the inbox. New agents inherit the company's mission and can themselves request reports.

Is this overkill for a single project?

Yes. For a one–off site or feature, Claude Code is the right tool. Paperclip earns its keep when you want autonomous, ongoing work across multiple workstreams.

What happens if the VPS goes down?

Auto–backups bring the company state back. Any task that was in flight needs to be re–run on the next heartbeat. Treat work as resumable, not as the source of truth — always have agents commit results to git or a real store.

Paperclip Tutorial: Multi-Agent AI Company

Introduction

For most of 2025 the conversation was about individual AI agents. The conversation in 2026 is about AI companies — full organisations of agents with a CEO, an engineering team, content roles, and reviewers, all coordinating through a shared dashboard, issues, and an org chart. The framework powering that pattern is Paperclip, an open–source orchestration layer that wraps around your existing LLM stack (Claude Code, Codex, Gemini, OpenClaw) and runs persistent agent teams that talk to each other on a heartbeat.

This is a complete setup guide: what Paperclip is, why running it on a VPS is the only sensible option, how the CEO heartbeat actually works, how to wire your subscription instead of an API key, how to write a CEO prompt that’s proactive instead of reactive, and how to structure agent teams with proper review loops so the work that comes out is something you’d actually ship.

📚 Table of contents

What Paperclip actually is
The org–chart demo — what a five–agent company looks like
Why you should not run Paperclip on your laptop
One–click VPS deploy on Hostinger
Signing in: API key vs. subscription (Claude / Codex)
SSH’ing into the Docker container to authenticate
Creating the company — mission, first agent, working directory
Issues, the heartbeat, and the Kanban states that matter
Writing a CEO prompt that’s actually proactive
Designing agent sub–teams with review loops
Cost controls, security, and connecting external tools
Common mistakes & pro tips
Frequently asked questions

🧠 What Paperclip actually is

Paperclip is an open–source orchestration platform on top of the agents you already use. It doesn’t replace Claude Code, Codex, or OpenClaw — it conducts them. You bring the model; Paperclip gives you a dashboard, an org chart, persistent agent sessions, an issue tracker, a heartbeat scheduler, and the glue that lets agents talk to each other rather than chatting only with you.

Stars on GitHub have climbed into the tens of thousands and the project sits at the front of a new wave of agent–company tooling. If you’ve ever wanted to point a team of AI workers at a goal and walk away, this is the shape that pattern is taking in 2026.

🏢 What a working Paperclip company looks like

A typical setup after a couple of days looks something like this: one CEO, one founding engineer, and a handful of role–specific agents (newsletter researcher, reviewer, writer). The only one you actually create is the CEO. Every other agent gets hired by the CEO once you approve the request — and those agents can hire their own reports.

📋 What the company gets done overnight

Analyses every previous newsletter and writes an improvement report.
Opens a PR against the main website repo with content optimisations.
Spins up its own GitHub repo to prototype a new newsletter platform.
Runs the CEO heartbeat every 10 minutes, picking up whatever is in the queue.

The mental model: you act as the board of directors. You set goals. The CEO delegates. Once the structure is in place, you wake up, check the inbox, approve a couple of hire requests or review items, and the company keeps going.

🖥️ Why you should not run Paperclip on your laptop

Paperclip can run locally. It really shouldn’t. The reasons are blunt and stack up fast.

⚠️ The risk surface

Multiple agents with root–level access to your filesystem.
Persistent network access — they can purchase, sign up, hit any URL.
A laptop that needs to stay on 24/7.
A single hardware failure wipes the whole company.

✅ Why a VPS solves it

Agents run inside a Docker container, isolated from anything important.
Uptime is the provider’s problem, not yours.
Auto–backups are a checkbox.
Starts at roughly $7/month if you skip local LLMs.

🚀 One–click VPS deploy on Hostinger

Hostinger ships a one–click Paperclip template that drops the project into a Docker container, exposes the dashboard over HTTPS, and pre–wires environment variables for the credentials you need to set.

Pick a KVM plan. KVM2 is a reasonable starting point if you plan to run more than a few agents.
Choose your billing period (12+ months unlocks partner discounts where available) and a region.
Enable daily auto backups.
Press Deploy from site, choose the Paperclip template, and let provisioning finish.

Once the instance is live you land on a configuration page. The fields that matter:

Admin name — default admin is fine.
Admin email — the login email for the dashboard.
Admin password — auto–generate a strong one and save it. This URL ends up on the public internet.
One model credential — an Anthropic, OpenAI, Gemini, or Cursor API key. Required to deploy, but you can remove it later if you prefer subscription mode.

🔐 API key vs. subscription — which to use

Paperclip can drive agents in two ways. The cost and the convenience curves are very different.

🔑 API key

Top up balance on the provider, paste the key into the Paperclip dashboard, set a monthly budget. Pay per token. Easiest path. Predictable.

🧾 Subscription (Claude Max, Codex, Gemini)

Sign into the CLI inside the Docker container and let Paperclip use your existing plan’s usage. Much cheaper for heavy parallelism. Check each provider’s ToS for headless usage before going down this path.

🛠️ SSH into the container to authenticate Claude Code or Codex

For subscription mode you have to log into your CLI from inside the Docker container so its session survives across heartbeats.

From the VPS overview, change the root password to something you know and copy it.
Open a local terminal and run ssh root@<your-vps-ip>. Type yes to trust the host, paste the password (your terminal won’t show keystrokes — that’s normal).
Run docker ps and copy the Paperclip container ID.
Run docker exec -it <container-id> /bin/bash to enter the container.
For Claude Code: type claude, follow the prompt, pick Claude account with subscription, copy the URL into a local browser, paste the returned code back.
For Codex: use codex login --device for headless device flow, then complete the authentication in a browser.
Type /exit when you’re back at a working agent prompt. Authentication persists.

🏗️ Create the company

Back in the dashboard, you walk through three short steps.

Name and mission. Be specific but tight. “Grow Tech with Tim on YouTube by automating processes, reviewing content ideas, creating scripts, and reviewing previous performance” works. A 10,000–word brief does not.
First agent (CEO). Pick Claude Code or Codex as the adapter. Hit Test; you want a green pass before continuing.
Working directory. Create a fresh folder on the VPS (e.g. mkdir paperclip-work inside your home directory) and paste its full path here.

Press Open issue and Paperclip launches the company. You’ll immediately get a hire request from the CEO for the founding engineer — approve it.

💓 Issues, the heartbeat, and the Kanban states

Paperclip isn’t a chat app. You don’t talk to the CEO. You file issues, and the agents pick them up.

🫀 How the heartbeat works

Only the CEO has a heartbeat. Every other agent gets triggered by the CEO.
Default interval is one hour. Drop it to 600 seconds (10 minutes) under the CEO’s configuration if you want faster turnaround — at higher cost.
Each heartbeat: the CEO scans the pipeline, picks up tasks, and delegates them.
You can also click Run heartbeat manually any time.

The Kanban board has the states you’d expect — backlog, to do, in progress, in review, blocked, done. The two you actually need to watch:

Blocked — the agent needs you. Usually a missing credential or external setup. The ticket explains exactly what.
In review — the agent finished but wants approval. Drop a comment with approved and the next heartbeat moves it forward.

Want something done immediately? Create the issue, set priority to Critical, and assign it explicitly to the CEO (or any other agent). Direct assignment triggers work on the next tick instead of waiting in the queue.

🧭 Write a CEO prompt that’s proactive, not reactive

The single biggest difference between a Paperclip company that hums and one that goes idle is the CEO system prompt. By default the CEO will often look at an empty queue and shrug. Fix that by giving it an explicit per–heartbeat protocol.

Skeleton for a good CEO prompt

Phase 1 — Status check. Iterate over every agent. Flag idle ones.
Phase 2 — Review completed work. Approve, return for revision, or escalate to the board (you).
Phase 3 — Assign work. Find an open issue (or invent one from the goals) for every idle agent. No agent goes a heartbeat empty–handed.
Phase 4 — Metrics & strategy. Read whatever analytics tools the agent has access to and adjust priorities.

Lazy shortcut that works: ask Claude itself to write the CEO prompt from a short brief. Paste in your company mission, list the agents, ask for “a strict per–heartbeat protocol that keeps every agent productive,” iterate until it reads like a real ops manual.

👥 Designing sub–teams with review loops

Single agents drift. Pairs help. Triplets with a reviewer are where Paperclip starts producing work you’d actually ship.

🧪 Newsletter team

Researcher → Drafter → Editor → Reviewer. Each agent has a slim system prompt. Output flows in one direction; only the reviewer can ship.

🛠️ Engineering team

Founding engineer with a strict git workflow: never push to main, always open a PR, never bypass review. Pair it with a code–review agent.

🎨 Content team

Topic researcher → script writer → thumbnail designer → publishing reviewer. Each has output format rules and a definition of done.

Cap the company at 10–15 agents to start. More than that, you can’t monitor effectively and you’ll burn budget on shadow work.

🧰 Cost controls, security, and external tools

Set a monthly budget in the CEO cost panel if you’re on an API key. Subscription mode doesn’t need this.
GitHub — create a dedicated bot account, sign into it on the VPS, and grant only the repo permissions you trust. Never the main branch.
External APIs — provide credentials inside an issue and tell the agent which tool to call. ConvertKit, Google Analytics, YouTube, anything with an API.
Inbox & review queue — check both daily. Inbox is the CEO escalating to you; review is unfinished work that needs approval.

🧨 Common mistakes & pro tips

❌ Common mistakes

Treating Paperclip like a chat box. There’s no chat. Only issues.
Skipping the CEO system prompt — the company will go idle within a day.
Hiring 30 agents on day one. You can’t supervise them, and the cost spirals.
Running on a laptop. Don’t.
Pasting credentials in plaintext in public–facing issues. Use the configuration panel.

✅ Pro tips

Drop the heartbeat to 10 minutes during the first day of tuning, then raise it.
Always give the CEO a step–by–step heartbeat protocol.
Build review loops into every team — researcher / writer / reviewer beats one super–agent.
Use critical–priority direct assignment to nudge the system when you need fast turnaround.
Sleep on the first run. Most of the value shows up overnight.

🏁 Conclusion

Paperclip is early. The UX needs polish, the prompts need handcrafting, and the leap from agent to organisation is not yet plug–and–play. But the shape is right: persistent agent sessions, a heartbeat scheduler, an issue tracker as the API, and a CEO that owns delegation. With a clean VPS deploy, a precise mission, a proactive CEO prompt, and 10 well–scoped sub–agents organised into review loops, you can leave the laptop closed overnight and wake up to real work shipped.

The interesting thing isn’t any individual run. It’s the workflow change. Once you stop chatting and start filing issues, you’re not using AI to type faster — you’re operating a small company.

Explore More on DevShelf

OpenClaw — Tool Profile
The single-agent runtime that Paperclip delegates individual company roles to — full overview of skills, channels, and pricing.
Deploy OpenClaw to the Cloud
Same Hostinger VPS setup as Paperclip — essential reading before you deploy your AI company.

Paperclip Tutorial: Run an Open-Source AI Company With Multi-Agent Orchestration