Verdict. It works as a control plane for agent work I’m already doing and can supervise closely. The creator calls the “zero-human companies” tagline “a bit aspirational.”
what it is
Open-source orchestration for running a company of AI agents. It sits above the agent framework. You define agents with roles (CEO, engineer, QA, designer), give them goals, and they work through a ticketing system with approvals, comments and per-agent budgets.
A Node.js server with a React dashboard and Postgres (embedded PGlite for dev). MIT licensed. The site is paperclip.ing.
The model is board (you), then a CEO agent, then specialist agents. Agents never message each other directly. Delegation means creating a task for another agent, coordination happens in task comments, and status is the task state. Each task has one owner, so the task graph doubles as the audit trail.
Agents wake up four ways. A timer, a new assignment, an on-demand trigger from the UI or API, or a system automation.
It drives other agents through adapters.
| Adapter | Status (April 2026) |
|---|---|
| Claude Code | Primary, most mature |
| OpenAI Codex | Full support |
| OpenClaw | Gateway adapter |
| Gemini | Experimental |
| Cursor | Background mode |
| OpenCode | Multi-provider |
| HTTP webhook, shell process | Any external system |
Run it locally with npx paperclip@latest (localhost:3100) or with Docker Compose, which comes with Claude Code and the Codex CLI pre-installed.
how I use it
I ran a content pipeline on it.
what I found
As of 2026-04-05, about five weeks after launch.
- About 47.7K stars. One contributor had 1,152 of the commits, so the bus factor was one.
- 1,765 open issues.
- Crash recovery was manual. Stale tasks didn’t get reassigned. That was the top issue, with 148 comments.
- No per-agent access control. Any member could task any agent.
- No local LLM support yet. Ollama was the most-requested feature.
- Errors compound without human checkpoints. One documented incident had batch outreach hit 23 leads instead of 3.
In a live walkthrough with Greg Isenberg, the creator was candid about where it fits. It started as a way to manage 20 to 30 Claude Code windows over a weekend, and the real traction is managing AI inside existing businesses, with approval gates on. He called imported 100-plus-agent org charts “completely unproven,” because there are no evals for them. He also named installing skills from public registries as an unsolved security risk.