Putting AI agents on an org chart, each with a budget
Paperclip is a Node.js server with a React UI that hands work to AI agents and tracks what they do with it. Each agent holds a position in an org chart, with a manager, a budget and tasks filed as tickets. Paperclip wakes an agent when a task is assigned to it, when someone replies or when a routine fires, and the ticket keeps the conversation, the files produced and the history of every run. The code is MIT-licensed and published by Paperclip Labs, Inc., which is preparing a hosted version, Paperclip Cloud, for now behind a waitlist.
Paperclip does not build agents. Flowise and Dify are for assembling an agent or a chain of LLM calls on a canvas. Paperclip takes agents that already exist, such as Claude Code or Codex, and decides who works on what and how much they may spend. Its README puts it this way: if an agent like OpenClaw is an employee, Paperclip is the company.
The comparison of open source workflow automation tools sets Paperclip against agent builders and classic workflow tools.
Hard budget caps that pause an agent at 100%
- Nine agents supported out of the box: the adapter docs list Claude Code, Codex, Gemini CLI, Cursor, OpenCode, Pi, Hermes, Grok Build and Kimi Code as stable. One team can mix them, each with its own model.
- One ticket, one run: a task has a single assignee, and an execution lock stops two concurrent runs from claiming the same ticket.
- Budgets that stop work: per organization, agent or project. Paperclip warns at 80%; at 100% the agent is paused and keeps its tasks.
- Review before closing: review and approval stages can be required before a task is closed or an agent is hired, and config changes are versioned with rollback.
- Routines: recurring work fires on cron, a webhook or an API call, and every run opens a tracked ticket.
Bring your own agents; no login out of the box
You bring the agents. Each adapter needs its tool installed (the Claude Code, Codex, Gemini or Cursor CLI, and so on), plus the API keys or subscriptions behind it. Signing in with a Claude or Codex subscription also requires Python 3 on the host; the Docker image ships Python and both CLIs. The OpenClaw, shell process and HTTP adapters work but cannot be picked in the UI yet: they are set up through the API or an organization export.
Node.js 24.11 and Postgres. npx paperclipai@latest onboard sets up an instance where a single Node process runs an embedded Postgres and stores files on local disk. For production, the README says to point it at your own Postgres. The documentation gives no memory figure.
No login by default, telemetry on. The default local_trusted mode listens on loopback and trusts whoever opens the board, with no sign-in. Remote access calls for authenticated mode, for instance with --bind lan or --bind tailnet. Usage telemetry, anonymous according to the README, is on by default; PAPERCLIP_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 turns it off.
Budget stops rely on recorded spend. The README warns that cost reporting and work already in flight can delay a stop. Monthly budgets reset on the 1st of each month (UTC), and an agent paused for budget alone comes back on its own.
Tickets live in Paperclip. Plugging in Jira, Linear or Asana is on the roadmap, not shipped yet. The README is written for people steering several agents toward one goal, and its example is twenty Claude Code terminals open at once.
Seven months old, with breaking releases. The repository was created on March 2, 2026. GitHub lists 28 releases between March 9 and October 6, numbered by date since March 18. The latest, v2026.1005.0, brings three breaking changes (an @-mention no longer starts a run) and database migrations 0284 through 0293. With close to 98,000 stars, the repository has 3,393 open pull requests against 2,895 merged, and 2,812 open issues. One account authored 2,980 of the 4,853 commits on the main branch.