TL;DR: Paperclip AI is an MIT-licensed, self-hosted control plane that runs a team of AI coding and task agents — Claude Code, Codex, OpenClaw, Cursor, Gemini CLI, OpenCode, and others — as one “company” with an org chart, budgets, approvals, and scheduled routines. npx paperclipai@latest onboard --yes gets you a dashboard at localhost:3100 in a few minutes. It is free software with no SaaS tier; you pay only your own model API bills. Two things to know before you deploy: it needs Node.js 24.11+, and any version below 2026.416.0 carries an unauthenticated remote-code-execution hole rated CVSS 10 (CVE-2026-41679, patched April 2026).

	Paperclip AI	n8n	Flowise
What it orchestrates	External AI agents (Claude Code, Codex, OpenCode…)	Workflows: 400+ app nodes, LLM steps among them	LLM chains and chatbots, visual builder
License	MIT	Fair-code (Sustainable Use License)	Apache 2.0
Model inference	None built in — each agent brings its own model	Via LLM nodes (OpenAI-compatible, Ollama)	Via chain nodes (OpenAI-compatible, Ollama)
Runtime	Node.js 24.11+, embedded PostgreSQL	Node.js or Docker	Node.js or Docker
Current version	2026.1005.0 (Oct 5 2026, CalVer)	actively released	actively released
The catch	CVSS 10 RCE below v2026.416.0; telemetry on by default	Not OSI open source; commercial hosting restricted	Narrower scope: chat flows, not long-running agents

Honest take: If you already run two or more CLI coding agents and keep losing track of what they did and what it cost, Paperclip is the first tool that treats that as a management problem rather than a scripting problem — and MIT licensing means no rug to pull. If your “agents” are really just LLM steps inside automations, n8n or Flowise remains the simpler, better-fitting pick.

Paperclip went viral in mid-2026 with the pitch “a self-hosted AI workforce for your business”: roughly 53,000 GitHub stars in its first six weeks, about 74,000 by August, and 98.9k stars with 16.7k forks as of October 9, 2026. Hype like that usually hides a thin wrapper. This one is not a wrapper — but what it actually is gets misreported constantly, so that is where a review has to start.

What does Paperclip AI actually do?

Paperclip AI is an orchestration layer, not an agent and not an inference server. The project (github.com/paperclipai/paperclip, MIT © 2026 Paperclip Labs, Inc.) ships a single Node.js server with a React dashboard, and every AI agent connects to it as an external runtime. You bring the agents; Paperclip gives them structure:

• Org chart: agents get roles and reporting lines — a “researcher” reports to a “lead,” the lead escalates to you.
• Budgets: per-agent and per-org spend thresholds with alerts and automatic pauses. With API-metered agents this is the feature that pays for the setup time.
• Task tracking and review gates: work items move through assignment, execution, and approval; you can require human sign-off before anything ships.
• Routines and triggers: cron-style scheduled runs plus webhook/API triggers, with heartbeat-based wakeups for long-lived agents.
• Shared infrastructure: an MCP tool gateway, encrypted secrets storage that agents reference without ever copying values, activity logs, and multi-organization support.

The supported-agent list as of October 2026 covers OpenClaw, Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, Hermes, Grok Build, and Kimi Code, plus escape hatches for any custom process or HTTP endpoint via adapter packages. That list is the tell: Paperclip is aimed at people orchestrating coding and operations agents that already exist on their machines, not at people building chatbots from scratch.

The license check comes back clean. The repository states MIT with no CLA gate on contributions and no openly stated restrictions, and there is no managed cloud version or paid tier at all — self-hosting is the only way to run it. The business model risk of a future license flip (the Terraform/n8n pattern) can never be ruled out, but MIT means the current code stays forkable forever.

How do you install Paperclip AI?

One command, if your Node.js is new enough:

$ node --version
v24.11.1   # requires >= 24.11.0 — this is the #1 install failure

$ npx paperclipai@latest onboard --yes
# downloads the server, provisions an embedded PostgreSQL database,
# and opens the dashboard at http://localhost:3100

The embedded PostgreSQL is a genuine friction-remover — no database container to manage on day one. Three other paths exist:

• npx paperclipai test-drive spins up an isolated throwaway instance; feed it ANTHROPIC_API_KEY or OPENAI_API_KEY and it demos a working agent org without touching your real config.
• From source: git clone, pnpm install, pnpm dev (pnpm 9.15+ required). Building the native runner from source additionally needs a Rust toolchain, or you point PAPERCLIP_RUNNER_BINARY at a prebuilt one.
• Docker: the official image bundles the Claude and Codex CLIs, which makes it the fastest path to a working coding-agent org on a fresh VPS. An entry-level Vultr instance is enough for the control plane itself — the server is a dashboard and a queue, not an inference engine — though agents that compile code or run browsers will want more RAM on the same box.

The Node 24.11+ floor deserves its own warning. Debian stable and Ubuntu LTS ship far older Node; if you deploy on a typical home-lab box, plan on installing Node via nvm or the NodeSource repo before anything else, or just use Docker and skip the problem.

One more default worth changing: telemetry is on out of the box. Set PAPERCLIP_TELEMETRY_DISABLED=1 (or DO_NOT_TRACK=1) if you run it for privacy reasons — it is also auto-disabled when CI=true. Sentry and OpenTelemetry integrations exist but are opt-in.

Does Paperclip AI work with Ollama and local models?

Not directly — and this matters if you came here from the self-hosted-LLM side. Paperclip performs no model inference of its own and has no model-provider settings to point at an Ollama endpoint. Models are configured per agent, inside each agent runtime: the README’s own examples use ANTHROPIC_API_KEY, OPENAI_API_KEY, and OPENROUTER_API_KEY, or local Claude/Codex subscription sign-in via the provider’s CLI.

So the local-model story is inherited from whichever agents you enroll. OpenCode, for example, talks to any OpenAI-compatible endpoint, which includes Ollama’s /v1 API — enroll OpenCode as a Paperclip agent and point it at your local server, and you have a locally-inferenced workforce member. Agents that only speak to their vendor’s cloud (Codex, Cursor) stay cloud-bound no matter what Paperclip does. A fully air-gapped deployment is therefore possible in principle but depends entirely on your agent choices, and you lose the bundled convenience paths. If local-first is the goal, budget VRAM for the agents’ models the usual way — a 24GB card covers the 27-32B coding models most local agent setups run (hardware math at runaihome.com) — and treat Paperclip purely as the coordination layer, which costs you nothing but CPU.

For picking which coding agents are worth enrolling in the first place, our sister site’s Kilo Code vs OpenCode vs Cline comparison covers the BYOK landscape.

Is Paperclip AI safe to put on your network?

Only on a patched version, and only deliberately. On April 9, 2026, CVE-2026-41679 was published: an unauthenticated attacker could chain six API calls into full remote code execution on any network-reachable Paperclip instance running in authenticated mode with default configuration — no credentials, no user interaction, CVSS 3.1 base score 10.0. The fix landed in version 2026.416.0; advisories conflict on whether 2026.410.0 already contained a partial fix, so treat 2026.416.0 as the floor. Anything current is fine — the project uses calendar versioning and the latest release is 2026.1005.0 — but a Paperclip box set up in Q1 2026 and never updated is remotely ownable today.

This hit an orchestrator in the worst possible place. A Paperclip server holds your encrypted secrets, your agents’ tool permissions via the MCP gateway, and — because agents execute real processes — a straight line to shell access on every enrolled machine. An RCE here is not “attacker reads my dashboard,” it is “attacker owns my agent fleet.” We made the same point about inference-layer exposure in the Hugging Face breach checklist and about agent containment after the Kimi K3 sandbox escape: the control plane is the crown jewels.

The defaults, post-patch, are sane: quickstart binds to loopback only (“trusted local mode”), and remote access requires explicitly enabling authenticated mode with a private-network bind — --bind lan or --bind tailnet, the latter designed for Tailscale. The practical rule: loopback or tailnet, never a public interface, and check npm view paperclipai version against your running instance monthly. The project’s 2.9k open issues against 4,900+ commits reads like an honestly triaged backlog for a repo this popular, not abandonment.

Paperclip AI vs n8n vs Flowise: which one do you need?

Ask what the unit of work is. If it is a workflow — data moves from app A through an LLM step to app B — n8n wins and Paperclip is the wrong shape entirely: n8n has 400+ integration nodes, mature error handling, and native Ollama support (our n8n + Ollama guide is the local-first setup). If it is a conversation — a RAG chatbot over your docs — Flowise’s visual chain builder gets you there fastest, Apache 2.0 and all; the full three-way with LangGraph is in Flowise vs n8n vs LangGraph.

Paperclip’s unit of work is an employee-shaped task: “research X and file a report for review,” “triage today’s issues, fix the trivial ones, escalate the rest.” Long-running, semi-autonomous, needs a budget cap and an approval gate. Before Paperclip, people stitched this out of cron, tmux, and spreadsheets. The “workforce” framing is marketing, but the primitives underneath — reporting lines, spend pauses, review gates — are exactly what multi-agent setups were missing. The honest counterpoint: if you run a single agent a few times a week, Paperclip is pure overhead; a shell alias is your orchestrator.

When NOT to use Paperclip AI

• You want a visual automation builder. Paperclip has no drag-and-drop workflow canvas. n8n or Flowise, full stop.
• You need strict local-only inference with zero cloud paths. Possible, but you are limited to agents that support OpenAI-compatible endpoints, and you must disable telemetry yourself. A plain Ollama + Open WebUI stack has a shorter audit surface.
• You cannot run Node 24.11+ or you refuse to track updates. The CVE history shows what a stale instance costs here.
• One agent, occasional use. The org chart, budgets, and approval machinery only pay off at two-plus agents doing recurring work.
• Production business process with an SLA. Paperclip is 2026-born software iterating fast under CalVer; treat it as a power tool for your own operations, not something you resell uptime on.

Verdict

Paperclip AI earns its stars. It is the first MIT-licensed tool that treats a fleet of AI agents as something to manage — with budgets, hierarchy, and audit trails — rather than something to script, and the embedded-PostgreSQL onboard is the smoothest first-run in this category. It is not a local-AI tool in itself: inference lives in your agents, privacy depends on which ones you enroll, and the April 2026 CVSS 10 CVE is a permanent reminder that the control plane must never face the open internet.

Your situation	Pick	Cost	Where
2+ CLI agents (Claude Code, Codex…) doing recurring work	Paperclip AI 2026.1005.0	$0 + your model API bills	npx paperclipai@latest onboard --yes
App-to-app automations with LLM steps, local models	n8n + Ollama	$0 self-hosted	n8n + Ollama guide
RAG chatbot over your own docs	Flowise	$0 self-hosted	Flowise vs n8n vs LangGraph
Always-on control plane off your home network	Paperclip on a small VPS	a few $/month	Vultr

FAQ

Is Paperclip AI really free, or is there a catch? The software is MIT-licensed with no paid tier, no cloud offering, and no feature gating as of October 2026. The real costs are your model API usage (each enrolled agent bills its own provider) and the server you run it on. The non-money catch: telemetry defaults to on, and you own all operational security.

Can Paperclip AI run fully offline with Ollama? Paperclip itself has no Ollama integration — it does no inference. You get local models by enrolling agents that support OpenAI-compatible endpoints (OpenCode being the clearest example) and pointing them at Ollama’s /v1 API. Set PAPERCLIP_TELEMETRY_DISABLED=1 and keep the server on loopback or a tailnet for a genuinely private setup.

Is the CVE-2026-41679 vulnerability still a risk in 2026? Only on unpatched instances. The unauthenticated RCE (CVSS 10) affects versions below 2026.416.0 and was fixed in April 2026; the current release is 2026.1005.0. If you installed Paperclip before April 2026 and never updated, patch before the instance touches any network.

Sources

• Paperclip AI GitHub repository — license, README, architecture, agent list (accessed Oct 9 2026)
• paperclipai on npm — version 2026.1005.0, Node >=24.11.0 engines field
• NVD: CVE-2026-41679 — unauthenticated RCE, CVSS 10, fixed in 2026.416.0
• CSO Online: Critical Paperclip bugs expose AI agent trust failures
• Paperclip project site — open-source statement, star count claims

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe