Paperclip: run your AI agents like a company
Someone gave nine AI agents job titles on Paperclip and had a live website three days later
- Who
- Published under MIT by Paperclip Labs, Inc. For people and teams running several coding agents at once.
- When
- The repository opened on 2 March 2026; the latest release is v2026.1023.0 (2026-10-24).
- Where
- A self-hosted Node.js server and web UI. Hires Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Hermes, OpenClaw and more.
What was hard
With twenty Claude Code tabs open, you soon lose track of which one is doing what. A reboot wipes that context, and you end up gathering it from several places to remind each agent what it was doing and why. Sometimes you only learn the next morning that a loop spent the night burning tokens.
What it does about it
Paperclip gives each agent a title, a manager and a budget on an org chart, and hands out work as tickets. Each ticket carries its project and company goal, so the agent gets the why along with the what. Agents don't run all the time: they wake when work is assigned, a comment arrives or a schedule fires, do one round of work and stop. When a budget hits its limit, that work pauses.

What people made with it
Nine Claude agents with roles (CEO, developer, QC, SEO, content writer) coordinated through Paperclip put a WordPress site live in three days; the CEO agent opened about 15 tasks on its own.
it even created a task telling me i was the blocker because i hadnt chased up some approvals. it was right

A one-founder company that has run more than 13,000 tasks on Paperclip turned the rules it wrote against loops and stuck work into three free MIT skills.
We run CyberNative AI LLC on Paperclip: AI agents do the work, and one human founder sets direction.
A read-only dashboard that polls Paperclip's API and shows each agent as a 3D character, with who is busy, workstream progress and a live activity feed.
I built an independent, read-only dashboard for Paperclip companies and wanted to share it.
How it runs inside
- 1
You (the board) → Paperclip server
Assign a ticket to an agentAssignment queues a wakeup for the assignee; tickets in backlog wake nobody.server/src/services/issue-assignment-wakeup.ts - 2
Routine (cron, webhook) → Paperclip server
Create a ticket on scheduleEach routine run creates an assigned ticket and wakes the agent through the same path.server/src/services/routines.ts - 3
Paperclip server → Claude Code
claude --print --output-format stream-jsonStarts Claude Code with a short-lived run token in PAPERCLIP_API_KEY, resuming a saved session.packages/adapters/claude-local/src/server/execute.ts - 4
paperclip skill → Claude Code
The heartbeat procedureOn waking: check assignments, check out, work, report, exit. It never runs continuously.skills/paperclip/SKILL.md - 5
Claude Code → Paperclip server
POST /api/issues/{issueId}/checkoutOne ticket, one agent: if another agent holds it, the call returns 409 and the agent moves on.server/src/services/issues.ts - 6
Paperclip server → Claude Code
Stop the run at the budget limitPast a company, agent or project budget, the server records a stop request on the run and halts it.server/src/services/heartbeat/budget-cancellation.ts - 7
Claude Code → Ticket comments and documents
PUT /api/issues/{issueId}/documents/{key}Results stay on the ticket as documents and comments, with run logs and usage records alongside.skills/paperclip/SKILL.md
What goes in, what comes out
Is it for you
Good for
- You keep several Claude Code or Codex windows open and lose track of who is doing what
- You want agents to take recurring work such as reports, support or social posts
- Your team wants agent spend held to company, agent and project budgets
- You want every agent result to pass a review or approval step
Not for
- It doesn't build agents: you bring your own (Claude Code, Codex and others) and pay for their models.
- Budgets stop work from recorded spend, so late usage reports or in-flight work can delay a stop (per the README).
- Anonymous usage telemetry is on by default; PAPERCLIP_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 turns it off.
- Agent Chat and the chat and email connectors are experimental, and a desktop app is only on the roadmap.
Reviewed by Devalley’s editor before publishing. How the column is made