A woman monitors a dashboard of AI assistants, task checklists, analytics, and security controls in a modern office.
Paperclip is an open-source tool that turns a set of AI coding agents into a pretend company. You act as the board and the agents act as staff. How-To Geek's Dibakar Ghosh spent a couple of weeks with it and came away a convert. Before you copy the setup, you should know what Paperclip does, what it costs to run, and what the security record looks like.

A woman monitors a dashboard of AI assistants, task checklists, analytics, and security controls in a modern office. What Paperclip is, and what it isn't​

Ghosh calls Paperclip a "meta-harness". A harness is the software wrapped around a language model that lets it use tools, loop, keep history and manage memory. ChatGPT, Claude Code and Codex are harnesses, and so are OpenClaw and Hermes. Paperclip sits above them. You connect whole harnesses to it, not individual models. It adds no new AI capability. It gives you a dashboard for assigning work, and it is not a chat window.

Paperclip's own documentation draws the same line. An agent there is a configuration: a role, an AI system, a budget and wake-up rules. The AI itself, whether Claude, Codex or something else, runs outside Paperclip. Independent write-ups agree. Paperclip ships no model of its own. The software is free, and you pay the agents' vendors for the agents.

So "free" describes the app, not the bill. You still need a Claude or ChatGPT subscription or an API key behind each agent. Paperclip's setup docs say as much.

The author's setup: Alfred & Co.​

Ghosh did not build a software company, which is the usual Paperclip demo. He built a personal concierge firm. These are his own configuration choices, not Paperclip defaults.

  • Alfred, the CEO. He runs on Codex through a ChatGPT subscription. For each new type of request, Alfred hires two or three agents, compares their answers and returns the best one. He fires the weaker agents and keeps the winner for next time.
  • A news-checking routine. Two Codex agents, both on a model Ghosh calls GPT Terra, hunt for breaking news. One looks for high-stakes stories. The other cross-references them against other sources.
  • Role matching. Ghosh says Claude suits planning, Codex suits coding, Grok suits X data research, and Gemini Antigravity or Codex suit design. These are his impressions, not benchmarks.

He also says he felt it was a gimmick at first, and that about a week in it had streamlined his workflow. His claim of winning back "a few precious hours" a day is his experience. He gives no measurements. The same goes for the cross-checking routine. Two agents agreeing is a process, not proof that a story is true. Review anything high-stakes yourself.

The most useful idea: an agent bake-off​

Ghosh says nobody is talking about this use. Many people give their most important task to the most powerful and most expensive model because they can't be bothered to test cheaper ones. With Paperclip, you can set up an "HR" agent that hires several "intern" agents. Each intern gets the same prompt on a different model, or slightly different prompts if you want to tune those too.

The aim isn't the best output. It's the cheapest combination of model and prompt that gives a good-enough result. Ghosh says one round can find a cheaper way to do the same job. It can also stretch a subscription's token allowance.

Treat this as a sensible experiment, not a proven saving. He offers no numbers. You still have to define "good enough" before you start, or you will just pick whichever answer reads nicely. Every intern also costs tokens, so the test only pays off for tasks you repeat.

Setting it up​

Ghosh gives this command:

npx paperclipai@latest onboard --yes

Paperclip's current quickstart and homepage show npx paperclipai onboard --yes, without the @latest tag. Use whichever form the live docs show. Here is the sequence:

  • Install prerequisites. Node.js is needed for the npx route. If you use Codex via ChatGPT or Claude Code, those CLI tools must be installed on the same machine. Paperclip's docs also mention Python 3 for local subscription sign-in.
  • Run the onboarding command. The quickstart says an embedded PostgreSQL database is used by default, so you don't set up a database yourself. The quickstart now defaults to trusted local loopback mode.
  • Open the dashboard. The local address is.
  • Create a company and define its goal. Then create the CEO agent. The docs say the first agent in any company is always the CEO.
  • Pick an adapter and model, then run "Test your agent". The docs say to wait for "Connection successful". If it fails, check the API key for stray spaces or sign in again. Also confirm the CLI opens outside Paperclip.
  • Go easy on heartbeats. Timed heartbeats start off. The docs warn that every heartbeat costs money and suggest hourly for a CEO at first.
  • Set budgets. The setup flow doesn't set a spending cap. The docs say agent and company budgets are separate, and either one pauses an agent when it runs out. They suggest $30–50 a month as a starting point for a CEO. That is project guidance, not a general cost estimate.

Ghosh says the whole thing takes about an hour. Treat that as his estimate, not a promise. Requirements change, and the project's own documentation is the authority.

Security: read this before you give agents your accounts​

The review doesn't cover security, and it matters here. These agents can run terminal commands, touch your files and use connected accounts.

Paperclip's documentation defines three deployment modes.

ModeLoginIntended use
local_trustedNoneOne person on one machine, bound to localhost
authenticated + privateRequiredTailscale, VPN or LAN
authenticated + publicRequired, with explicit public URLInternet-facing deployments with stricter checks

The default localhost:3100 setup has no login. It is fine for solo experiments. Don't expose that port to the internet.

The project's GitHub advisory page lists several serious issues. The titles include:

  • Critical: unauthenticated remote code execution through an import authorization bypass.
  • Critical: drive-by remote code execution against local instances via DNS rebinding.
  • Critical: cross-tenant agent API token minting.
  • High: an agent that inherited a connected Gmail account and could send real email.
  • High: malicious skills able to exfiltrate and destroy user data. This one is dated April 16, 2026; the July 22, 2026 date belongs to the DNS-rebinding advisory above, and the rest of the list is April 2026.

One independent write-up counts 12 advisories: 5 critical, 4 high and 3 medium. I haven't checked which versions are affected or fixed, so I can't say whether the current release is safe. Read the individual advisories and update before use. Another independent analysis warns that third-party skills run with the same permissions as your agents, often meaning full filesystem and network access. Audit skills before installing them.

Practical takeaways​

  • Stay in local mode while you experiment.
  • Use authenticated mode if you need access from other devices.
  • Set budget caps before enabling heartbeats.
  • Keep production credentials and personal accounts away from experimental agents.
  • Run the bake-off on one repeatable task before you hand your best model's job to a cheaper one.

This isn't a Windows-specific story. The documentation I reviewed doesn't establish Windows compatibility details, so check the live docs for your platform. The core idea is useful whatever you run. A dashboard for agents can save management effort, but it also multiplies cost and risk unless you keep limits on both.


Correction (November 21, 2026): The malicious-skills advisory was listed as Critical and dated July 22, 2026. GitHub lists it as High, published April 16, 2026; July 22, 2026 is the publication date of the DNS-rebinding advisory.

 

References

  1. I built my own company of AI agents using Paperclip, and I regret not doing it sooner How-To Geek 2026-11-21T15:30:15+00:00
  2. your-first-agent.md github.com
  3. github.com github.com