paperclipai/paperclip

The problem

Once you run more than a handful of agents—Claude Code here, Codex there, an OpenClaw bot on a cron job—you lose the thread. Who is working on what, which task gated which release, how much this month's token spend actually was, and whether the agent that just touched production had permission to. Chat windows and terminal tabs don't answer those questions. Paperclip puts a task-manager-shaped UI over a multi-agent operation: assign goals, define roles and budgets, approve work, and read the audit trail in one place.

What's different

Most agent tooling solves a single-agent loop or a single-vendor runtime; frameworks like LangGraph or CrewAI give you code-level orchestration primitives, and a hosted IDE agent keeps one session coherent. Paperclip instead treats each agent as an employee in a structure—org chart, scoped secrets, approval gates, cost tracking—and is deliberately provider-agnostic, bridging Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Hermes through adapters. The interesting trade-off is that it does not try to be the model harness; it is the management layer above whatever harness you already run. That makes the governance and budgeting pieces the real product, not the prompting.

Who it's for

Small teams and solo operators already running several coding agents daily who need visibility, spend limits and review checkpoints—especially anyone coordinating mixed human and agent work. It is not for people running a single agent in an editor, nor for teams that need a hardened enterprise control plane today: the README is heavy on positioning and screenshots, thin on architecture, security model and failure behavior.

Cost of adoption

Node.js server plus a React UI, self-hosted, MIT licensed, so no vendor lock-in but you own the deployment and the secrets. Every supported harness needs its own credentials and adapter setup, and the whole model costs money per agent per month—budgeting is a feature because the bill is real. The project is young (created March 2026) and the README does not document install commands; you will be following the docs site. Expect churn in the adapter surface as vendor CLIs move.

Verdict

Paperclip is the most explicitly "agent company" framing in a crowded field, and the org-chart-plus-budget angle is genuinely different from code-level frameworks. The signal is loud: roughly 107k stars and ~17.7k forks in about eight months, with +25.5k stars in the last tracked month alone (#4 on the monthly board)—that is a spike, not a slow burn, and it landed without daily trending appearances or much Hacker News discussion, which usually means distribution did the work rather than peer review. Treat it as promising and unproven: pilot it on a bounded internal workflow where you can tolerate adapter breakage, and keep your own audit logs until the governance claims are validated in practice.