Paperclip — the company around the agents

paperclipai/paperclip · MIT · TypeScript · 75 090★ / 13 977 forks · created 2026-03-02, pushed daily since. Author: the pseudonymous @dotta; the repo now sits under Paperclip Labs with 100+ contributors. Docs at docs.paperclip.ing.

Their one-liner is the whole thesis: “If OpenClaw is an employee, Paperclip is the company.”

It is not another agent runtime. It’s a Node.js server + React UI that sits above the agents you already run and gives them the thing scripts can’t: roles, reporting lines, budgets, approval gates, and an audit trail. As the author put it — “you can only manage a rats nest of shell scripts and HEARTBEATS.md for so long before you realize there’s got to be a better way.”

The workflow

You hire a CEO agent; it drafts a hiring plan and delegates to specialists. You sit above all of it as the board — approving hires, budgets, and anything you’ve flagged as needing sign-off. Delegation then flows up and down the org chart on its own.

Bring your own agent

Adapters for OpenClaw, Claude Code, Codex, Cursor, bash CLI agents, and HTTP/webhook bots, plus external adapter plugins. The admission rule: “If it can receive a heartbeat, it’s hired.”

This is the part that makes it interesting for us — it composes with the runtime already on the machine instead of replacing it. The README’s own qualifying signal is uncomfortably familiar: “You have 20 simultaneous Claude Code terminals open and lose track of what everyone is doing.”

The four pillars

Twelve server-side subsystems back this: Identity & Access, Work & Tasks, Heartbeat Execution, Governance & Approvals, Org Chart & Agents, Workspaces & Runtime, Plugins, Budget & Costs, Routines & Schedules, Secrets & Storage, Activity & Events, Company Portability.

What to steal

The engineering claims are more valuable than the branding. Five mechanics worth lifting regardless of whether we ever run Paperclip:

1. Atomic task checkout + budget enforcement in one transaction. No double-work when several agents wake at once, no runaway spend racing the check. This is the correct answer to the failure mode parallel orchestrators mostly leave to the human.

2. Per-agent monthly budget that hard-stops. Not a warning — the agent stops. “Runaway loops waste hundreds of dollars of tokens and max your quota before you even know what happened” is a real failure we’ve hit; a hard ceiling per agent is the fix, and it belongs in our own tooling.

3. Goal ancestry carried on every task. Tasks hold company → project → goal → parent links, so an agent always sees the why, not just a title. Our skills pass artifacts forward but never the chain of intent — this is a concrete upgrade to the pipeline.

4. DB-backed heartbeat queue with coalescing, plus persistent agent state across heartbeats and automatic recovery of orphaned runs. Compare with the cron→Workers→Prefect ladder: this is rung two done properly, and it beats a pile of systemd timers.

5. Portable company templates. Export/import orgs, agents, and skills with secret scrubbing and collision handling — the same idea as our stack templates, but for an entire org definition. Also: config changes are revisioned and bad ones roll back.

Runner-up worth noting: isolated execution workspaces built on git worktrees and operator branches, with dev servers and preview URLs attached — the worktree-per-agent pattern, but as infrastructure rather than a GUI convenience.

Where the hype ends

The “zero-human company” framing is marketing. In practice a Paperclip company always has a human at the top: you’re the board, approving hires and budgets while agents work in between. Zero busywork is accurate; zero humans is not.

Independent testing lands harder. One reviewer built a company over a weekend and got “an AI-generated brand, a marketing strategy full of hallucinated statistics, and a website that looked like someone fed a blender a stack of Bootstrap templates.” AI automation specialist Nick Puru calls the broader pattern “productivity theater” — the seductive belief that organizing tasks equals achieving results. Hallucinated numbers presented as fact, unreviewed code reaching production, security and logic errors: all still on the human.

The honest read: a real architectural trend (org-level orchestration is genuinely missing) wrapped in overpromising branding. The org chart solves coordination, not quality — quality still needs evals in CI and review gates, which is exactly why their fourth pillar bothers to include evals and agent performance reviews.

Why it matters for a solo builder

Two things, neither of which is “fire yourself”:

• Multi-company isolation on one deployment — every entity is company-scoped with separate data and audit trails. That’s a portfolio of small bets with one control plane, which maps directly onto how we run several projects at once.
• Cost as a first-class primitive. Budgets, throttling, and cost events per agent turn “how much did this experiment burn?” into a number on a dashboard instead of a surprise invoice — relevant to kill/iterate/scale thresholds, which need per-project cost to be decidable.

Install is a single command (npx paperclipai onboard --yes, Node 20+, pnpm 9.15+), self-host only, no account required, two deployment modes (trusted local or authenticated). A “Clipmart” template marketplace — content agencies, trading desks, dev shops — is announced but not shipped.

Links

• paperclipai/paperclip — the repo (verify the org: lookalikes exist)
• docs.paperclip.ing — docs
• agent-toolkit-landscape — where it sits in the ecosystem
• conductor-parallel-agents — the layer below: many agents in worktrees, no org
• project-solo-factory — our pipeline: skills as the unit, not roles
• background-jobs-ladder — heartbeats vs the scheduling ladder
• agent-self-discipline — limits and drift, the manual version of budget enforcement
• harness-engineering-summary — approval gates and audit as harness, not prompt