Paperclip, the app people use to manage AI agents for work

My notes

Summary

Paperclip is an open-source, self-hosted app that orchestrates a team of AI agents as a “company”, org chart, goals, budgets, governance, rather than as standalone tools. It is runtime-agnostic: any agent (Claude Code, Cursor, Codex, scripts, webhooks) that can receive a heartbeat can be “hired” and pointed at one mission. MIT-licensed, embedded Postgres for local runs, no vendor account required.

Key Insight

• Mental-model shift, not another tool. Framing moves from “I am prompting an AI” to “I am managing a team.” You act as board of directors; the CEO agent proposes strategy, you approve, agents execute.
• Bring-your-own-agent. No lock-in, adapters connect to Claude Code sessions, OpenClaw bots, Python scripts, shell commands, or HTTP webhooks. “If it can receive a heartbeat, it’s hired.”
• Hard cost ceilings per agent. Monthly budget per agent; soft warning at 80%, auto-pause + task block at 100%. Costs tracked per agent / task / project / goal. Directly addresses runaway-token risk.
• Heartbeat scheduling, not always-on. Agents wake on schedule or notification (assigned ticket, @-mention), check work, act. Delegation flows up/down the org chart automatically.
• Audit-first. Every instruction, response, and tool call is a ticket with full tracing and an append-only immutable log. Governance gates (e.g. hiring a new agent needs board approval) are default-on.
• Multi-tenant. One deployment can run dozens of isolated “companies”, useful for parallel ventures or templating org configs.
• Uses a SKILL.md convention so agents discover the context they need, same pattern as Claude Code skills.