Paperclip AI is an open source orchestrator that converts Claude Code, Codex and Cursor into a single team, with organizational hierarchy, per-agent budgets and approvals. You install it on your own infrastructure, it’s MIT licensed and doesn’t require a Paperclip account. Its latest stable version is v2026.916.0, released on September 16 with 503 commits.
As of the publication date of this note (September 28, 2026), the repository had accumulated 91.9K stars and 15.8K forks, and that same day added 3.2K stars on GitHub Trending.
What is Paperclip AI?
Paperclip (also searched as Paperclip AI) is a Node.js server with a React interface that coordinates a team of AI agents toward a business objective. The README sums it up in one line: if OpenClaw is an employee, Paperclip is the company.
It looks like a task manager. Underneath, it models a complete organization:
- A company with an objective. The quickstart example is building a notes app until reaching USD 1M of MRR.
- Agents with roles. A CEO, a CTO, engineers and marketers, each with a position, direct manager, permissions and budget.
- Tasks that carry their entire chain of objectives, so the agent sees why they’re doing something, not just the ticket title.
- Heartbeats. Agents wake up according to a schedule or events (an assignment, an @-mention), review their work and act.
The typical first use: you create a company, hire a CEO agent and approve the strategy they propose. From there they start creating tasks on the board.
The README is equally explicit about what Paperclip isn’t:
- it’s not a chatbot
- it’s not an agent framework
- it’s not a drag-and-drop workflow builder
- it’s not a prompt manager
- it’s not a code review tool
It doesn’t build agents. It manages the organization where your agents work, what in English is known as multi-agent orchestration or AI agent orchestration.
What agents can you connect to Paperclip?
Paperclip works with any agent that can “receive a heartbeat”. The official adapter reference currently lists:
- Claude Code
- Codex
- Gemini CLI
- Cursor (local and cloud)
- OpenCode
- Pi
- Hermes (local and gateway)
- Grok
- Kimi Code
- OpenClaw Gateway
It also has generic Process and HTTP adapters, plus an SDK for writing your own. Sandbox execution is supported via providers like e2b, Cloudflare, Daytona, Modal and self-hosted Kubernetes.
If you read our coverage of OpenClaw or Hermes Agent, this is the layer that goes above them. Hermes is an integrated adapter since version v2026.626.0.
Paperclip or OpenClaw: which do you need?
You don’t have to choose, because Paperclip uses OpenClaw, Claude Code and company. The project’s FAQ answers it directly: Claude Code or OpenClaw do the work, and Paperclip turns several of them into an organization with budgets, objectives, governance and accountability.
The README draws the line clearly. With a single agent, you probably don’t need Paperclip. With twenty, you do.
Its “problems it solves” table reads like the diary of anyone seriously using agents:
- twenty Claude Code tabs open not knowing what each one does
- sessions lost on restart
- context you have to explain by hand every time
- an uncontrolled loop that burns through your daily token budget before anyone notices
It also answers the obvious question of why not just connect your agents to Trello or Asana. Coordination has subtleties that a generic manager doesn’t solve:
- who has a task checked out
- how sessions persist
- how spending limits are applied
Paperclip solves task checkout and budget control atomically, so two agents can’t work the same ticket at once. If you’ve already tried a lighter approach like oh-my-claudecode, which orchestrates within Claude Code, Paperclip plays at a different scale: multiple providers, multiple people and multiple companies in the same instance.
How do you control the cost of a multi-agent system?
Cost control is where Paperclip earns its place for whoever signs the invoice:
- Monthly budget per agent. When an agent reaches their limit, it stops.
- Warnings before cutoff. Budget policies can warn when crossing a threshold, before final cutoff.
- Overages are contained. If the limit is exceeded, agents are paused and queued work is automatically canceled.
- Detailed tracking. Token consumption and costs are broken down by company, agent, project, objective, issue, provider and model.
The official documentation provides its own estimate of what you’ll pay your model provider:
- between USD 5 and 20 to try out the product
- between USD 20 and 100 per month for an active company
These are figures from the provider itself, not measured by us. The documentation also recommends defining budgets per agent and per company before activating heartbeats.
Is Paperclip AI free?
Yes, the software is free and open source (MIT). You host it yourself and don’t need a Paperclip account.
What you pay for is model usage: your Anthropic or OpenAI API key, or your existing Claude or Codex subscription. Since version v2026.916.0, those credentials can be connected via the new Connections system.
The release notes also mention a hosted Paperclip Cloud. As of the publication date of this note, we couldn’t find public pricing for that service.
How to install Paperclip AI?
You need Node.js 24.11 or higher. The recommended path downloads the installer and verifies its checksum before running it:
curl -fsSLO https://paperclip.ing/install.sh
curl -fsSLO https://paperclip.ing/install.sh.sha256
if command -v sha256sum >/dev/null 2>&1; then
sha256sum -c install.sh.sha256
else
shasum -a 256 -c install.sh.sha256
fi
bash install.sh
The installer:
- ensures you have a compatible version of Node.js
- installs a managed CLI
paperclipaiin~/.paperclip/cli - starts the interactive onboarding
- optionally installs Paperclip as a background service on compatible Linux and macOS systems
To try it without installing anything permanently:
npx --registry https://registry.npmjs.org paperclipai onboard --yes
That path starts in trusted local mode (loopback), the fastest way to get a first run. To use it in authenticated mode within your network or tailnet, explicitly choose a bind preset:
paperclipai onboard --yes --bind lan
# or:
paperclipai onboard --yes --bind tailnet
From source code you need Node.js 24.11+ and pnpm 9.15+:
git clone https://github.com/paperclipai/paperclip.git
cd paperclip
pnpm install
pnpm dev
This spins up the API server at http://localhost:3100, with an embedded PostgreSQL database that’s created automatically. For production, you point it to your own Postgres.
To use the Claude Code adapter, Claude Code has to be installed on the same machine.
Does Paperclip work on Windows?
Partially, and it depends on which piece we’re talking about.
- Local agents run natively on Windows. Since version v2026.722.0, Claude, Codex, Gemini, and custom ACP agents run natively on Windows in addition to Linux.
- The managed CLI installation and service lifecycle are documented for Linux, macOS, containers, and WSL.
- The install script is Bash. On Windows, the realistic paths are WSL or the path using
npxwith Node.js 24.11+ installed.
What’s in version v2026.916.0?
The September 16th version focuses on credentials and scope.
Already available:
- AI credentials move to Connections. Claude and Codex subscriptions, API keys, and shared accounts become managed accounts with permissions. They follow the responsible person during agent hiring and task execution. The same agent can run with one person’s subscription and another person’s API key.
- GitHub access moves to per-person identities. Identities backed by a GitHub App replace a single shared token.
Experimental (disabled by default or behind a flag):
- AgentMail inboxes for agents
- native chat connectors for Slack, Discord, Telegram, and Microsoft Teams, plus iMessage via Photon
- Agent Chat within the app
- the Paperclip Runner execution engine
Breaking changes worth knowing about:
- The “cheap model profiles” execution mode is removed.
X-Forwarded-Hostis now only respected if it comes from a trusted proxy. If you run Paperclip behind a reverse proxy, confirm thatTRUST_PROXYis configured.
What should a security team review before deploying it?
There are two key points for anyone putting Paperclip on a shared server.
- Update to v2026.916.0. Before this version, agent API endpoints returned stored credentials in plain text to any call with permission to read the agent, including the agent itself (PR #9860). The new version passes all responses through a redaction layer, and the project itself describes the previous behavior as a leak.
- Telemetry is enabled by default. The project claims it doesn’t send personal information, prompts, file paths, or secrets, and that it hashes references to private repositories. You can disable it with
PAPERCLIP_TELEMETRY_DISABLED=1orDO_NOT_TRACK=1. It disables automatically whenCI=true.
On governance, the controls a CTO would expect are already there:
- approval gates
- versioned configuration with rollback
- secrets scoped per agent
- secret proposals that a human must approve
- an immutable activity log that attributes each change to an actor
What’s still not available?
At the time of publishing this note, the public roadmap marks as not yet delivered:
- Memory/Knowledge
- a desktop app (documentation points to an unofficial community one)
- work queues
- bring your own ticketing system (Asana, Linear, or Jira as entry point)
- Connected Apps with one click
Cloud deployments are listed as partially delivered.
Is Paperclip AI worth it?
If you’re running a single agent, Paperclip is overkill. If you’ve already reached the point where multiple agents run in parallel across different providers, it gives you the pieces you’d otherwise build poorly on your own:
- a task board that agents can’t book twice
- budgets with automatic cutoff
- approvals
- an audit log
It’s the most complete open source answer we’ve seen so far to a question we already asked ourselves: does orchestration matter more than the model?