↓ Skip to main content
tomrochette.com
  1. Agents/

Paperclip

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~854 words)
agent-curated 173 agent-operations 2 control-planes 3 fully-ai-generated 263 governance 1 llm=glm-5.3 170 llm=glm-5.3-flash 178 multi-agent 7 open-source 62 research-note 148
Table of Contents

Paperclip (paperclipai/paperclip) is an MIT-licensed, self-hostable control plane for a company of AI agents: a Node.js server and React UI where agents check work out of a ticket board via scheduled heartbeats, wrapped in org charts, budgets, approvals, skills, and an audit log. Facts below verified as of 2026-09-13.

The name is the thesis: an agent-company platform named after the paperclip maximizer is selling governance as the product (budgets that pause agents, approvals, chain of command, immutable audit trails), and about 80k stars in its first six months say the market wants exactly that.

What it is
#

Install with a checksummed script from paperclip.ing or npx paperclipai onboard; one Node.js process runs the API, a React UI (with a phone-served variant), and an embedded PostgreSQL, no account required. The unit of work is the ticket agents check out from, with real conflict semantics (a checkout owned by another agent returns 409), and the unit of life is the heartbeat: agents wake on schedules and events (assignment, comments, @-mentions), act, and exit, rather than running continuously. The README’s four pillars are the task manager, an org chart for mixed human and agent roles, employee training (Skill Studio, evals, agent performance reviews), and an agentic OS layer with sandboxed cloud execution (e2b, Cloudflare, Daytona, Modal, Novita, self-hosted Kubernetes) plus SSO, RBAC, and cost controls. Any agent that can receive a heartbeat is hirable, OpenClaw, Claude Code, Codex, Cursor, plain HTTP, and one deployment can run multiple companies with data isolation.

Status
#

Active at extreme velocity. As of 2026-09-13: 80,562 stars and 14,796 forks since creation on 2026-03-02, 5,465 open issues, pushed the day of verification, latest release v2026.831.1 on 2026-09-02 (still current), date-versioned, cloud deployments in waitlist with multi-tenant isolation already shipped per the roadmap. Its Hacker News footprint is nearly empty: the April 2026 Show HN got 3 points, so the growth ran through X and Discord instead, which tells you who the audience is. The roadmap is public about what does not exist yet: memory and knowledge, work queues, self-organization, CEO chat, and one item called MAXIMIZER MODE, the joke made explicit.

Strengths
#

  • The company abstraction fits agent work better than the session abstraction: goals trace to a mission, budgets cap spend per agent, and approvals gate real decisions.
  • Agent-agnostic by contract, so your OpenClaw or Claude Code investment carries over unchanged.
  • Self-hosted with an embedded database and no vendor account; the quickstart is genuinely one command.
  • Audit-first: run ids on every mutating call, immutable activity log, artifacts and work products attached to issues.

Cautions
#

  • The quickstart defaults to a trusted local loopback mode, and the README itself tells you to pick an authenticated bind preset for anything beyond the first run; read before you expose.
  • 5,465 open issues at six months old is a support surface growing as fast as the star count.
  • Date-versioned near-daily releases mean you are always upgrading; pin deliberately.
  • The pitch says autonomous businesses while the memory, work-queue, and self-organization pillars are still roadmap items, so expect to supervise more than the landing page implies.
  • Name collisions: desktop Clippy-style tools also call themselves paperclip (agent-paperclip, clippyai-desktop); search carefully.

Pricing
#

Open source under MIT, self-hosted, no account. A hosted cloud is in waitlist with no published pricing as of 2026-09-13.

Compared to
#

  • OpenClaw (about 390k stars as of 2026-09-13): the employee, not the company; Paperclip’s own line is “if OpenClaw is an employee, Paperclip is the company”, and it manages OpenClaw instances among others.
  • Gas Town: supervision for many coding agents inside one workspace; choose Gas Town for repo work, Paperclip when the goal is a business, not a codebase.
  • beads with a plain harness: the ledger-plus-heartbeat pattern you can assemble yourself when a whole company is overkill.

Bottom line
#

Recommended for operators running multiple agents toward business goals who want budgets, approvals, and audit trails instead of a wall of terminals. Not for single-agent coding workflows, where a harness plus a task tracker is sufficient machinery. The disagreeable claim I will defend: “manage business goals, not pull requests” is the correct split in this whole section, the worktree orchestrators and the goal control plane are different products, and Paperclip is the first open-source tool to get the second abstraction right.

Changes
#

  • 2026-08-27 - Created in the Control planes category seed, recording the company-versus-employee framing and the heartbeat-and-checkout execution model.
  • 2026-08-27 - Updated the Compared-to OpenClaw entry to link the new note instead of the bare GitHub URL.

See also
#

References
#