Short answer: On 6 November 2026 (16 Aban) we hired / created a 3-agent AI company — Researcher, Writer, and Reviewer — in a separate local SOP Paperclip install, with periodic heartbeats off and a $1 company hard-stop plus $0.25 per-agent hard-stops on Paperclip billed_cents. Only Researcher actually ran — about 4 minutes and 12 seconds, using Claude via a subscription — over public Sophrai and competitor sites. Writer and Reviewer were created and configured but had no runs. The research returned an internal n8n migration/audit + email-checklist idea (target ≥16 of 200 visitors in 14 days) with clear caveats. No commercial results or revenue were claimed or produced; nothing was published or sent. Paperclip recorded $0 billed (subscription_included). The provider-reported equivalent API value was about $1.50 (1.496797 USD). The Paperclip billed_cents budget stayed at 0 and did not cap subscription usage. That is the $0 catch.
Paperclip is the orchestration layer (company, agents, tasks, budgets). The model/subscription is separate. Sources: official GitHub and docs linked below. Pilot figures are from our run receipts, not a vendor marketing claim.
Hands-on pilot note (6 Nov 2026, Asia/Tehran): this was a separate local SOP company for testing. We are not publishing local dashboard URLs. No affiliate links for Paperclip in this article.
Title / video: I Hired a 3-Agent AI Company (Paperclip): Real Run, Real Cost, the $0 Catch. The Persian counterpart (this article stays English) is «یه شرکت 3 نفرهی هوش مصنوعی ساختم (Paperclip): اجرای واقعی، هزینهی واقعی و نکتهی صورتحساب صفر». This article is the written companion. The YouTube embed above stays a placeholder until the public video ID is known. Channel: sophraiAI on YouTube.
What Paperclip is (and is not)
Paperclip is open-source orchestration for teams of AI agents. The README’s framing is useful: it looks like a task manager; underneath you get org charts, budgets, governance, and coordination. You bring adapters/models (for example Claude Code or Codex local adapters per the docs). Paperclip is not the model. The model provider still runs inference; Paperclip schedules agents, tracks work, and enforces policies on the costs it records as billable.
Official entry points:
- Repository: github.com/paperclipai/paperclip (MIT)
- Product site: paperclip.ing
- Docs home: docs.paperclip.ing
- Five-minute path after install: Quickstart: Your First Agent in 5 Minutes
- Costs & budgets: Costs and Budgets · Set a monthly budget
- CLI install: Installing the Paperclip CLI
What we actually did on 6 November 2026 (16 Aban): only Researcher ran
This was a controlled pilot in a separate local-only SOP company — not a production change to any live client site, and not an automated “replace your employees” demo.
- Company + three agents. We hired / created a 3-agent AI company — Researcher, Writer, and Reviewer — in a separate local SOP Paperclip install, wired for a short research → draft → review path.
- Periodic heartbeats off. We did not leave a 24/7 loop running. Work was intentional: one research invocation, not continuous heartbeats.
- Budgets before the run. Company monthly hard-stop at $1.00 (
budgetMonthlyCents: 100/ policy amount 100 cents). Per-agent hard-stops at $0.25 each. Policies used metricbilled_cents,windowKind: calendar_month_utc,warnPercent: 80,hardStopEnabled: true,unpricedUsagePolicy: block. - Only Researcher ran. About 4 minutes and 12 seconds on 6 November 2026 (16 Aban), using Claude via a subscription. Public Sophrai pages and public competitor sites only. Goal: find a concrete, shippable offer idea with a measurable early success bar.
- Writer and Reviewer had no runs. They were created and configured, then left idle. We did not start a draft or review pass.
- No publication, no outreach, no commercial result. The research output stayed internal. We claimed no revenue and produced none. We did not publish content, send emails, or change other WordPress pages from this pilot.
Writer and Reviewer had no runs
Creating a 3-agent AI company is not the same as running one. Researcher completed a single job (~4 minutes and 12 seconds). Writer and Reviewer never executed, so this pilot is not evidence that a full research → draft → review company worked end to end, and it is not a claim that we replaced a task list with a functioning three-agent operation.
What the research produced (internal only)
The researcher’s draft recommendation was roughly: ship a browser-side n8n workflow auditor / pre-migration checklist (no upload required for the first cut) aimed at freelancers and small agencies, capture email, then sell migration help. Suggested early success metric: ≥16 of the first 200 visitors complete the tool + email in 14 days. Explicit caveats in that draft: third-party “version 3” claims still need primary-source verification, and USDT-only checkout can reduce conversion. Treat that as a hypothesis from one run — not a validated product plan.
The $0 catch: Paperclip billed $0, subscription usage still happened
| Signal | What we saw |
|---|---|
| Date | 6 November 2026 (16 Aban) |
| Who ran | Researcher only — about 4 min 12 sec, Claude via subscription |
| Writer / Reviewer | Created and configured; 0 runs |
| Commercial result / revenue | None claimed or produced |
| Paperclip billed / cost cents | 0 ($0 billed) |
| Provider final receipt | subscription_included, cost status reported / provider_reported |
Provider-reported equivalent API value (costUsd) | ≈ $1.50 (1.496797 USD) |
Paperclip billed_cents budget | Stayed at 0 and did not cap subscription usage — the $0 catch |
| Tokens (provider receipt) | ~64,863 input · ~764,418 cached · ~14,156 output |
So the honest headline is not “this 3-agent AI company cost $0” or “we replaced our task list with a working three-agent staff.” It is: we hired a 3-agent company, only Researcher ran, Paperclip recorded $0 billed (subscription_included), the provider-reported equivalent API value was about $1.50 (1.496797 USD), and the Paperclip billed_cents budget stayed at 0 and did not cap subscription usage. That is the $0 catch.
That matches how the official docs separate concerns: budgets pause scopes when recorded billable spend hits the cap; subscription seats and quota windows are tracked in Costs/Subscriptions and with the provider. If your adapter marks usage as included in a subscription, expecting billed_cents alone to act like a dollar kill-switch for all real-world usage is the wrong mental model.
We stopped further Paperclip model runs after seeing that gap. This article does not claim a guaranteed actual-dollar cost for repeating the pilot.
Boundaries (read these before you copy the setup)
- Local-only SOP company. Reproducible in spirit from public docs; our instance URLs and internal IDs are not public assets.
- Heartbeats off by default in the pilot. Official docs warn that agents make real provider calls and recommend budgets before enabling heartbeats. We agree — and would add: know whether your adapter bills API cents or subscription quota.
- One Researcher run, not a self-running 3-agent company. Writer and Reviewer had no runs. Delegation and CEO strategy approval are powerful; they are also how spend multiplies if left unsupervised.
- No screenshots fabricated for this draft. When the matching YouTube video is public, the video is the visual walkthrough.
- No affiliate pitch for Paperclip here. Links go to official GitHub/docs only.
- Research claims ≠ shipped product. The n8n auditor idea was internal output from the pilot, not a Sophrai release announcement tied to this run.
Public reproducible overview (quickstart distinction)
Use the public install path from the README/docs — do not invent a private installer:
# Requires Node.js 24.11+ per current docs/README
npx paperclipai@latest onboard --yes
paperclipai run
Then, after the instance is up (docs “five-minute path”):
- Create a company and set a clear goal.
- Hire/configure the first agent (often a CEO in the guided path) with an adapter such as
claude_localorcodex_local— that adapter is how the model runs; Paperclip orchestrates it. - Approve the first strategy if the CEO proposes one.
- Before enabling periodic heartbeats: set company and per-agent monthly budgets (UI under Audit → Budgets, or the budget API/how-to). Defaults in docs: warn ~80%, hard-stop at 100% of
billed_cents, calendar month UTC. - Add specialist agents only when you have a real task graph. For our pilot shape: Researcher → Writer → Reviewer, with heartbeats left off until you intentionally wake a run.
Orchestration vs model, in one line: Paperclip decides who works when and under which budget; the model provider does the thinking and meters tokens/subscription quota.
Who this suits
- You already run multiple Claude Code / Codex / similar sessions and lose track of ownership.
- You want approvals, org reporting lines, and a task board around agents — not another chat tab.
- You can read provider receipts and Paperclip Cost/Budget views side by side.
Who should wait:
- Anyone hoping “company budget = $0 forever on a subscription.”
- Teams that need a managed SaaS with no self-hosting (Paperclip’s public pitch is open-source / self-hosted).
- Operators who will enable heartbeats without a kill-switch and without watching subscription quota.
What we would do differently next
- Treat provider subscription quota and Paperclip
billed_centsas two different brakes; configure both. - Keep heartbeats off until a single run’s receipt is understood end-to-end.
- Write the success metric and “do not publish” rule into the company goal before the first wakeup.
- Pair any public article with a real video walkthrough — hence the placeholder embed above.
Sources used for this article
- Paperclip GitHub repository (README positioning, MIT license, quickstart command)
- Paperclip Docs · Five-minute path · CLI installation
- Costs and Budgets guide · Set a monthly budget how-to
- Pilot operational receipts from our 6 Nov 2026 (16 Aban) local SOP run: only Researcher ran (~4 min 12 sec, Claude via subscription); Writer and Reviewer had no runs; subscription_included; Paperclip billed $0 / billed_cents stayed 0 and did not cap subscription usage; provider costUsd≈1.496797; token counts above
FAQ
What is Paperclip, exactly?
Paperclip is an open-source, self-hosted control plane for AI agent teams: companies, org charts, tasks, approvals, heartbeats, and budgets. It is not itself the language model. Official positioning: if a single agent is an employee, Paperclip is the company. See the GitHub README and docs.paperclip.ing.
Did all three agents run?
No. We hired a 3-agent AI company (Researcher, Writer, Reviewer) in a separate local SOP Paperclip install, but only Researcher actually ran — about 4 minutes and 12 seconds on 6 November 2026 (16 Aban), using Claude via a subscription. Writer and Reviewer were created and configured but had no runs. No commercial results or revenue were claimed or produced.
Did the $1 Paperclip budget stop all spending?
No. In our pilot the Paperclip ledger stayed at billed_cents=0 / costCents=0 because the adapter reported subscription_included usage. The provider receipt still showed ~$1.50 equivalent costUsd with tens of thousands of tokens. Paperclip budgets enforce billed_cents hard-stops; they do not automatically replace your provider’s subscription quota limits.
What did the research agent recommend?
An internal draft only: a browser-side n8n workflow audit / pre-migration checklist plus email capture, with a success bar of at least 16 of the first 200 visitors completing the tool + email in 14 days. Caveats included third-party version claims needing verification and USDT-only payment friction. We did not ship that product or any outreach from this run.
How do I try Paperclip without copying our internal setup?
Use the public path: Node.js 24+, npx paperclipai@latest onboard –yes, then paperclipai run (or the docs five-minute path: create a company, hire a CEO agent, approve strategy). Set company and per-agent budgets before enabling periodic heartbeats. Full steps: docs.paperclip.ing.
Who is this for?
Builders who already juggle several coding/research agents and want org structure, approvals, and spend controls in one place. It is less useful if you only need a single chat window or a simple task list with no multi-agent coordination.