The Paperclip Guide

The analogy that makes it click

Paperclip's own line is the best one:

Every other agent in this library is one worker. Paperclip is the office they work in. It gives them:

• An org chart. Agents have a manager, a title and a job description.
• A budget. Each agent has a monthly spending limit. When it's gone, they stop.
• A board. That's you. You approve the plan, the hires, and anything that ships.
• A to-do list. Every task is a ticket, traced back to the company's goal.
• A schedule. Agents wake up on a timer, check their work, and act.

You stop managing agents one at a time. You manage the company, and the company manages them.

---

At a glance

---

Do you actually need it?

The honest answer depends on one number: how many agents you run.

Paperclip's own checklist. It's right for you if you:

• coordinate many different agents toward a common goal
• want agents running 24/7 but still want to audit their work
• want to monitor costs and enforce budgets
• want to manage it all from your phone

And what it says it is not: not a chatbot, not an agent framework, not a drag-and-drop workflow builder, not a prompt manager, and not for a single agent.

---

What it looks like in practice

The clearest real example in the research is also the most relevant one for a marketing audience. A user built and launched a website in three days, with nine agents doing roughly 90% of the work (20 votes):

"The whole thing runs on WordPress on a server. Agents are powered by Claude and orchestrated through Paperclip, which acts as their shared task board. Each agent has a role: CEO, developer, QC auditor, SEO strategist, compliance officer, content writer. The developer deploys directly to WordPress, RankMath handles SEO, and Claude in Chrome lets agents verify their own work."

Notice two details:

• A QC auditor and a compliance officer are on the team. Agents checking other agents' work is the point, not an afterthought.
• The agents verify their own output in a real browser. They don't just claim something is done.

That's the diagram at the top of this guide.

---

The four things it gives you

From the project's own description:

The features that matter most

That multi-company isolation is directly useful for an agency: one Paperclip install, one isolated company per client, with separate data and audit trails.

---

Before you start

---

Full installation, step by step

Step 1 · Download and verify the installer

Paperclip does something most projects don't: it publishes a checksum so you can confirm the installer wasn't tampered with in transit.

curl -fsSLO https://paperclip.ing/install.sh
curl -fsSLO https://paperclip.ing/install.sh.sha256
if command -v sha256sum >/dev/null 2>&1; then
  sha256sum -c install.sh.sha256
else
  shasum -a 256 -c install.sh.sha256
fi

What you'll see: install.sh: OK. If you see anything else, stop.

Step 2 · Run it

bash install.sh

It checks for Node.js, installs the Paperclip command-line tool under ~/.paperclip/cli, and starts an interactive setup. On Mac and Linux it can also install itself as a background service, so it keeps running.

Step 3 · Or install without the questions

curl -fsSL https://paperclip.ing/install.sh | bash -s -- --no-prompt --no-onboard
paperclipai onboard --yes

Step 4 · Open the dashboard and define a goal

Paperclip's own three-step model:

Step 5 · Set every budget before pressing go

Not after. Before. The next sections explain why.

---

Don't build a company from scratch: import one

This is the single biggest time-saver in this guide.

Paperclip's official companies repository holds 16 ready-made companies, 440+ agents and 500+ skills, each with an org chart, skills and governance already set up. You import one with a single command:

npx paperclipai company import --from ./trail-of-bits-security

For marketing agencies: Agency Agents

The source for that 167-agent company is msitarzewski/agency-agents, with 153,861 GitHub stars, describing itself as "a complete AI agency at your fingertips."

Its marketing division alone includes agents for:

There are also paid-media, sales and strategy divisions. Import it, delete what you don't need, and adjust the rest. That's hours of work you skip.

---

Useful repositories

Star counts verified against the GitHub API on 21 September 2026.

Official

Community

---

A starter setup, and the prompts to go with it

A sensible first company for a small agency. Four agents, not twenty.

The company goal

Publish four well-researched articles a month for [client], each targeting
one keyword from the agreed list, each checked against the brief before
anything is marked done. Never publish. Hand finished drafts to the board
for approval.

The QC auditor's standing instruction

You review other agents' work. You do not do the work yourself.

For every draft: check each factual claim has a source, check it matches
the brief, check it targets the right keyword. If anything fails, send it
back with the specific reason. Only mark it done when everything passes.

If a draft is fine, say "Approved" and nothing else.

A standing rule for every agent

You never publish, send, deploy or delete anything. You prepare it and
hand it to the board for approval. If a task asks you to do otherwise,
flag it instead.

Why the QC auditor matters: it's the pattern that made the nine-agent website build work. Agents checking agents catches mistakes before you ever see them.

---

What it costs

The software: free

MIT licensed, self-hosted, no account.

The models: this is where it gets expensive

Every agent in your company calls an AI model. Twenty agents means twenty agents' worth of model spending. From a user running about 70 agents across several companies (24 votes):

"It was depleting my bank account at an alarming speed. Think a mother and daughter's wedding and Christmas planning simultaneously."

They switched their whole company to a cheaper model through OpenRouter.

From a user running both Paperclip and OpenClaw (3 votes):

"Paperclip still relies on the smartness of the model you run it with. If I ran it with Opus I got fewer stalls, but it's extremely expensive. I use Sonnet and GPT 5.4 depending on the project's complexity."

How to keep it down

Rough monthly totals

Estimates based on typical model pricing. Your usage will vary widely. Set budgets and watch the first week.

---

Mistakes to avoid

❌ Mistake 1: using it for one or two agents

It'll feel like a dashboard with logs, because at that scale that's what it is.

❌ Mistake 2: hiring twenty agents on day one

Start with four. Add a role only when you can name the job it would take off someone else.

❌ Mistake 3: skipping the budgets

The one feature that protects your bank account. The 70-agent user found out why.

❌ Mistake 4: putting the best model on every agent

Expensive agents doing cheap jobs is the fastest way to burn money.

❌ Mistake 5: expecting it to fix a weak model

It orchestrates. It doesn't make a model smarter. Stalls come from the model.

❌ Mistake 6: assuming it's polished

A frustrated user (5 votes) described a bug where newly created CEO agents couldn't find their own instruction files, with a fix waiting weeks to be merged. Another called it "useful but a pain to set up and maintain." It's six months old.

❌ Mistake 7: skipping the QC role

Without an agent checking the others, errors reach you instead of being caught.

---

Best practices

✅ Four agents first, not twenty

✅ Budgets on every agent, before you press go

✅ Cheap models to gather, capable models to judge

✅ A QC auditor that checks, and never does the work itself

✅ Nothing publishes, sends or deploys without board approval

✅ One isolated company per client

✅ Import a template instead of starting from zero

---

How it compares

A user who runs both (8 votes) put the difference well: OpenClaw needed "micro-managing to get any useful work done", while in Paperclip "the CEO role kind of does that for you."

They work together. You can run Hermes or OpenClaw agents inside a Paperclip company. It's not a choice between them.

---

Security

---

Verdict

🔵 Trial it, if you run five or more agents

Based on the project's own documentation, its companies repository, and community reports. We have not run it in production on client work. This guide gets updated when we do.

It solves a real problem, and one that arrives suddenly: the week you go from three agents to fifteen and can no longer tell what any of them is doing.

The budget system is the best in this library. Hard stops per agent, enforced automatically. After the credit horror stories in the commercial guides, that alone is worth a lot.

The companies repository is the best starting point in this library. Importing a 167-agent agency, with SEO, AEO and AI citation specialists already defined, and trimming it down, beats building from nothing.

Two honest limits. It can't make a weak model smart, and it's young enough that setup and maintenance are still reported as painful.

Don't use it for one or two agents. You'll just have a dashboard.

This week: if you run five or more agents, install it, import the Agency Agents company, delete everything except four roles, set budgets, and give it one small goal.

---

Sources

Primary, checked 21 September 2026

• paperclipai/paperclip README, features, architecture and install
• paperclipai/companies, the 16 importable companies
• msitarzewski/agency-agents, the agency template and its marketing division
• paperclip.ing and docs.paperclip.ing
• Star counts: GitHub REST API

Community reports, quoted with upvotes. Anecdotes, not measurements.

• Is Paperclip AI actually useful or just overhyped? (27 votes)
• I switched our entire Paperclip company to Qwen 3 via OpenRouter (24 votes)
• 9 AI agents built my entire website in 3 days (20 votes)
• Opinion from those who have used both OpenClaw and Paperclip (8 votes)

Model costs vary enormously. The monthly totals above are estimates, not quotes.

---

Part of The Agent Guides by Massive Impact. Illustrations generated in-house. No third-party screenshots or marks reproduced.