Paperclip + OpenClaw: Running Agents as a Company

Bottom Line

Paperclip is not another agent. It is the layer above them.

If OpenClaw is one capable worker, Paperclip is the thing that gives five of them an org chart, a budget that stops at 100%, a schedule, and a paper trail. It launched March 4, 2026, is MIT-licensed, and passed 30,000 GitHub stars in three weeks — which tells you the problem it solves was widely felt.

The reason it belongs on this site: it is agent-agnostic. Claude Code, Codex, Cursor, OpenClaw and Hermes all plug in as workers. It sits on top of hardware you already own.

What it actually does

A Node.js server and a React UI, wrapping four ideas:

• Org chart. Agents have positions and can delegate work downward, rather than every task landing on one agent.
• Budgets that auto-pause at 100%. Each agent gets a monthly allowance and stops when it is gone.
• Scheduled heartbeats. Agents wake on a schedule and check whether there is work, instead of waiting to be prompted.
• Audited tickets. What was assigned, what was done, by which agent.

Its own framing is a “human control plane for AI labor”. The honest description is narrower and more useful: it is the accounting and scheduling layer that multi-agent setups keep reinventing badly.

Why the budget feature is the interesting one

Unmetered agent spend is the standard way these setups fail. An agent loops, or picks an expensive model for trivial work, and you find out from the invoice.

It is worth comparing directly to what shipped last month. Grok Bot’s early reviews centre almost entirely on cost control: no published weekly allowance, no spend cap, no model picker, with users reporting 42% of a week consumed on day one. Paperclip’s answer to the same problem is a per-agent budget that halts at 100%.

One is a hosted product that has not shipped the control yet. The other is an MIT-licensed project where the control is the headline feature.

If your workers run on local weights, the budget matters less — the marginal cost of a local run is electricity. It matters most for the mixed setup most people actually have: local models for the bulk work, a paid API for the hard calls.

Where OpenClaw fits

Paperclip handles hiring, budgeting and auditing. OpenClaw does the work.

That division is worth being precise about, because the two projects get compared as rivals and they are not. You do not choose between them any more than you choose between an employee and a payroll system.

Nous Research maintains an official hermes-paperclip-adapter that runs Hermes Agent as a managed employee inside a Paperclip company. The adapter spawns the Hermes CLI in single-query mode, Hermes processes the assigned task with its full tool suite, then exits. That is the pattern: the control plane assigns, the agent executes and exits, the ticket records it.

An OpenClaw worker follows the same shape.

When you do not need this

Most people reading this do not, yet.

One agent doing one job does not need an org chart. If you run a single OpenClaw instance handling your inbox or your repo, Paperclip adds a server, a UI and a scheduling concept for no gain.

The threshold is roughly: more than two agents, work that must be delegated between them, or spend you cannot currently see. Below that, it is overhead.

It is also young. March 2026, moving fast, and a star count is popularity rather than production evidence. Treat the audit trail as a feature you should verify against your own workload before trusting it for anything regulated.

The hardware question

Nothing about Paperclip changes what your rig needs — it schedules work, it does not run inference.

What it changes is how many agents you are realistically going to run at once, and that does affect hardware. Two agents sharing a 24GB card is a queue. Four agents on a schedule that wakes them together is a memory problem you will discover at 3am.

If you are moving from one agent to a team, size for concurrency rather than for the largest model. The gear ladder lists what each tier holds, and the 64GB guide is the usual starting point for running more than one model resident at a time.

What we have not tested

We have not run a Paperclip company in production. The description above comes from the project’s own documentation and published reviews.

What we can say from this side: the layer it occupies is real, most people building multi-agent setups end up writing a worse version of it, and the budget-with-a-hard-stop is the right default. Whether the audit trail holds up under a real workload is the thing to check before you depend on it.