Paperclip turns your AI agents into a self-running company

If you find yourself drowning in twenty open terminal windows running Claude Code or constantly jumping between Cursor and Codex, you know the pain: managing AI tools can quickly become more exhausting than doing the work yourself.

Open-source project Paperclip wants to solve that coordination mess by wrapping a Node.js server and React dashboard into a virtual company framework. The project's pitch is simple: if an agent like OpenClaw is an employee, Paperclip is the company.

Instead of micromanaging pull requests, you act as board chair. You give the system a top-level goal, such as building a note-taking tool that hits recurring revenue targets, and let an org chart of specialized bots break it down and execute.

An org chart powered by heartbeats

Under the hood, Paperclip behaves less like an experimental autonomous experiment and more like enterprise management software.

The system isolates businesses into distinct company units with separated data and budgets. Within each company sits an org chart where every agent has a role, permissions, and assigned models. It works with diverse runtimes including Claude Code, Cursor, Codex, OpenClaw, and Gemini CLI. As the team puts it: if it can receive a heartbeat, it can be hired.

Work moves through a structured ticketing system. When you define a mission, the AI CEO breaks it down into quarterly targets, which AI managers and engineers turn into individual tasks with clear goal alignment.

To keep costs and compute sane, agents do not run continuously. They wake up on scheduled heartbeats, atomically check out tasks using database locks to prevent duplicate work, execute the step, submit deliverables, and go back to sleep. Full state snapshots are captured at every step, allowing tasks to recover cleanly from crashes.

Hard budget limits and company templates

Letting autonomous bots run wild usually sounds like an easy way to burn through API credits overnight. Paperclip builds guardrails right into the runtime.

Budgets cascade from the company level down to departments and individual agents. Tasks run pre-execution cost checks, subtract spend upon completion, and trigger a hard freeze when limits are exceeded. Major inflection points like hiring new bots, altering corporate strategy, or initiating external payments require human approval.

Paperclip also lets users export their entire setup—roles, skills, and agent configs—into portable templates that resolve API key conflicts upon import. The maintainers plan to roll out Clipmart, a marketplace where users can share and clone complete AI company architectures.

Why it matters

Right now, calling this a self-running business is generous. Given the reality of hallucination and complex debugging, Paperclip feels closer to a detailed corporate simulation game than a money-printing machine. You will not hit a button and watch bots autonomously operate a profitable startup from scratch.

Even so, we think Paperclip makes a remarkably clever pivot away from agent hype and toward boring, necessary systems engineering.

Most multi-agent experiments fail because they lack accountability, state persistence, and cost control. Paperclip provides exactly those guardrails: heartbeat polling, atomic task checkouts, role-based access, and budget fuses.

For solo founders juggling multiple coding assistants, it offers a coherent operational framework right now. More importantly, it sketches out what real AI-native governance looks like well before the rest of the industry catches up.