I built my own company of AI agents using Paperclip, and I regret not doing it sooner

I built my own company of AI agents using Paperclip, and I regret not doing it sooner

Published Nov 21, 2026, 10:30 AM EST Dibakar Ghosh is a tech journalist at How-To Geek with over a decade of experience using and writing about open-source software, Linux, Android, AI, and personal productivity. He writes about software, operating systems, and tools that help people get more from the technology they already use. Most conversations about AI revolve around companies replacing their employees with it. While many people are scared by the prospect of losing their jobs, it also presents an interesting possibility: you could create your own AI-powered company, where you're the board, and everyone else is an AI agent. Now, if that sounds a bit too far-fetched, well, do I have news for you! A ton of AI tools have been popping up to help you build your own company of AI agents. And I've spent the past couple of weeks testing one such tool to see whether it's actually useful or just an utter gimmick. Meet Paperclip It’s not an AI tool itself, but a tool for managing your other AI tools Paperclip is a free, open-source app for running teams of AI agents. Technically, it's what you'd call a meta-harness — a harness for other AI harnesses. You see, an LLM (Large Language Model) on its own can only process your input, reason about it, and generate an output, whether it be text, image, music, or something else. A harness is the extra layer of software built around an LLM that allows it to use different tools, run execution loops, maintain conversation history, manage memory — basically, make the LLM useful and relevant. For example, ChatGPT is a harness built around OpenAI's GPT models. Likewise, Claude Opus and Claude Sonnet are the LLMs, while Claude Code and Claude Cowork are the harnesses. Popular AI tools like OpenClaw and Hermes are essentially LLM-agnostic harnesses that let you plug in different LLMs. Now, Paperclip is a management layer for connecting and managing all these different harnesses. Think of it as a central hub for OpenClaw, Hermes, Claude Code, OpenAI Codex, Google's Antigravity, and other agentic tools. You're not connecting individual LLMs to Paperclip — you're connecting entire harnesses. The idea isn't to introduce new AI capabilities to your LLMs, but to give you a management layer that helps you orchestrate the different AI harnesses you're already using. Paperclip Paperclip is a free, open-source harness for running your other AI agents, like Claude Code, ChatGPT Codex, OpenClaw, and Hermes. It offers a company-like paradigm where, instead of the traditional chat interface, you interact with your different AI agents by assigning them tasks. Each AI agent functions like an employee of the company, while you're the company director who has the final say. Paperclip is a better way to use your existing AI tools Paperclip isn't a chat-focused harness like Claude Code or Codex. It feels more like a company management dashboard. You essentially serve as the company's "board," assigning tasks to your employees and having the final say in how things are run. You can hire different agents with different abilities, and these agents essentially function as your company's employees. The idea is that every harness is good at some things and bad at others. So, when tackling a task, you can have multiple agents approach it from different directions. For example, I've found Claude particularly good at reasoning but expensive for coding. So Claude Code can serve as a senior developer writing the plan on how to approach a piece of software, while Codex can be the actual coder writing the code. Likewise, Grok has direct access to X (or Twitter) data, so you can use it as your market researcher. And Gemini Antigravity or Codex can serve as your graphic designer, depending on which model's artistic capabilities you prefer. You control how much access each agent has, its system prompt, and which specific LLM tier it uses. You can have high-agency employees, such as a CEO or HR agent, who report directly to you and can create (or hire) other agents as needed. You can also have low-agency agents handling the grunt work and reporting only to the CEO, so you're not flooded with unnecessary updates. Now, it might seem like Paperclip is only relevant if you use multiple AI harnesses, such as Claude Code, Hermes, and OpenClaw, together. And while that is indeed its primary appeal, it can be just as useful even if you're subscribed to only one AI harness. For example, you could potentially run a company entirely with Codex agents. The CEO and CTO could use more powerful models, such as GPT Astra or Sol, while less demanding roles, such as reporting or coding, could be handled by lighter models like GPT Terra or Luna. And of course, each agent's actual function is determined by its system prompt and the skills and connectors it has access to. The best part is finding the right agent for each job If all of this sounds a bit unnecessary, gimmicky, and like a solution in search of a problem, I get it. I felt exactly the same way when I decided to try Paperclip. However, after using it for about a week, I found that it actually helped streamline my workflow. Oh, and I also discovered a really cool use case that absolutely no one seems to be talking about. If you're anything like me, you might have an important task that needs doing, and you always assign the most powerful (read: expensive) model to that job because you don’t want to risk it. You also don't have time to A/B-test different AI models and benchmark their results to find the best price-to-performance ratio. Well, now with Paperclip, you can do exactly that by setting up an HR agent that hires multiple "intern" AI agents and puts them to the test to see which performs best. Each intern can use the same prompt paired with a different model, or you can give them slightly different prompts if you want to optimize the prompts as well. Of course, the agent running the most powerful model will do the best job, but we’re not necessarily looking for the best. The goal is to find the model-and-prompt combination that delivers a "good enough" result at the lowest possible cost. With a single round of testing, this method can help you identify a more cost-effective way to do the same task. And even if you're not using LLMs at API rates, you can use this to optimize token consumption and get more out of your subscription. How I used Paperclip to build my own concierge company I have an AI butler that hires and manages my other servants Most people I've seen using Paperclip so far are using it to run a software company. They've built systems where a senior developer agent running a powerful model works on the PRD (Product Requirements Document), a junior developer agent running a more affordable model handles the token-heavy work of writing code, another agent monitors GitHub for issues and fixes bugs, and so on. It's a legitimate use case, but I'm not looking to start a software company. I wanted to find something more personal that would still demonstrate what Paperclip can do. I eventually settled on creating a personal concierge company called Alfred & Co. Alfred is the CEO and my personal butler, powered by Codex through my ChatGPT subscription. I tell him what I need, and he hires the help. I've configured him so that whenever I request a new type of task, he'll hire two or three agents to tackle it, compare their answers, and give me the best result. He'll then terminate the weaker agents while keeping the best one around for the next time I need something similar done. In addition to this, I've also set up a routine in which two more Codex-powered agents, both using GPT Terra, research the web to find and fact-check breaking news. One agent focuses on finding potentially high-stakes news, while the other cross-references it against other sources to verify its accuracy. How to set up Paperclip It barely takes an hour The Paperclip developers plan to release it as a web app in the future, but at the time of writing, it's only available as a local app you need to self-host. That being said, it's ridiculously easy to get started. All you have to do is run this single command in your terminal: npx paperclipai@latest onboard --yes Paperclip requires you to have Node.js installed on your system. If you don't have it already, you can check its official site for instructions on how to install it. Paperclip will automatically set up the database and everything else, so you don't have to worry about the technical details. Once the installation is complete, you can access the dashboard at http://localhost:3100. From there, you can create a company, add your first agent, pick its harness, and connect the subscription or API key you want it to use. Do note that if you're using Codex through ChatGPT or Claude Code as the underlying agent, their respective CLI tools must also be installed on the same machine for Paperclip to access them. You can check out the official guide for details on using Paperclip, or simply feed it to your favorite AI harness for a more personalized walkthrough. This is my new favorite way of using AI moving forward Using AI as a chatbot is so 2024. The most productive way to use AI right now is through an agentic approach, where you hand it a task and let it do the work for you. I've been moving much of my workflow toward this approach, and while it does help streamline most of my admin chores, I still end up sitting at my desk for just as many hours micromanaging all the agents. However, with Paperclip, I can finally streamline the management side of things as well and win back a few precious hours every single day.

Original Source

Read the full article at Howtogeek →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.