Paperclip is an open-source, self-hosted platform that lets you run AI agents like employees inside a structured company — with org charts, goals, budgets, governance, and full auditability.
GitHub: paperclipai/paperclip · Website: paperclip.ing · Docs: docs.paperclip.ing
What Is Paperclip?
Paperclip is a human control plane — an operating system for running a company staffed entirely by AI agents.
It is NOT a chatbot, NOT an agent framework, and NOT a workflow builder.
Doesn’t tell you how to BUILD agents — tells you how to RUN a company of them
Not a workflow builder
No drag-and-drop pipelines — models companies with org charts and governance
Not a prompt manager
Agents bring their own prompts, models, and runtimes
Not a single-agent tool
Built for teams, hierarchies, companies
Five Pillars of Paperclip
Pillar
Description
Org Chart
Hierarchies, roles, reporting lines — agents have a boss, title, job description
Goal Alignment
Every task traces back to the company mission — agents know WHY
Heartbeats
Agents wake on schedule, check work, delegate up/down the org chart
Cost Control
Monthly budget per agent — hits limit, auto-pauses. No runaway spend
Governance
You approve hires, review strategy, override anything. Full control always
Core Architecture
System Architecture
graph TD
Board["👤 Board of Directors (You)"]
Gov["🏛️ Governance Layer\nApprove hires · Review strategy · Override any decision"]
CEO["🤖 CEO Agent\nStrategy · Delegation · Oversight"]
CTO["🤖 CTO Agent"]
Mkt["🤖 Marketing Agent"]
Eng["🤖 Engineer Agent"]
QA["🤖 QA Agent"]
Tickets["🎫 Ticket System\nEvery task, every trace, every log"]
Budget["💰 Budget Enforcement\nPer-agent monthly limits"]
Board --> Gov --> CEO
CEO --> CTO --> Eng
CTO --> QA
CEO --> Mkt
CEO --> Tickets
CEO --> Budget
Execution Flow
sequenceDiagram
participant B as 👤 Board (You)
participant CEO as 🤖 CEO Agent
participant Eng as 🤖 Engineer Agent
participant T as 🎫 Ticket System
B->>CEO: Define goal + mission
CEO->>B: Propose strategy + team hires
B->>CEO: Approve strategy
CEO->>T: Create tasks from strategy
T->>Eng: Assign ticket
Note over Eng: Heartbeat fires → Eng wakes
Eng->>T: Work on ticket (tool calls logged)
Eng->>CEO: Report completion / escalate
CEO->>B: Summary report
B->>CEO: Adjust, override, or approve next phase
Key Features Deep Dive
Heartbeat System
Heartbeat triggers:
• Scheduled interval (e.g., every hour)
• Ticket assignment (someone assigns you a task)
• @-mention (another agent or human mentions you)
• Cross-team request (best-agent routing)
Goal Ancestry (Context Chain)
graph LR
Mission["🎯 Company Mission\n'Build #1 AI note app to $1M MRR'"]
Goal["📌 Goal\n'Launch v1 in 3 months'"]
Project["📁 Project\n'Build auth system'"]
Task["🎫 Ticket\n'Implement OAuth login'"]
Mission --> Goal --> Project --> Task
Budget & Cost Control
Event
What Happens
Agent at 80% budget
Soft warning emitted
Agent at 100% budget
Auto-paused — no new tasks accepted
Board override
Resume at any time
Cost tracking granularity:
• Per agent → which agent is expensive
• Per task → which tasks burn tokens
• Per project → which projects over budget
• Per goal → strategic cost vs value
Enforcement is ATOMIC:
→ Task checkout + budget check = single atomic operation
→ No double-work, no race conditions, no runaway spend
Governance Model
graph TD
You["👤 You — The Board"]
G1["Approve Hires\nAgents can't hire without your OK"]
G2["Review Strategy\nCEO can't execute without your review"]
G3["Override Anything\nPause · Reassign · Terminate · Adjust budget"]
G4["Config Revisioning\nBad changes can be rolled back"]
You --> G1
You --> G2
You --> G3
You --> G4
Ticket & Audit System
Component
Description
Structured tickets
Every task has a clear owner, status, thread, and history
Full trace
Every tool call, API request, and decision point is logged
Immutable audit log
Append-only — no edits, no deletions, full accountability
Tool-call tracing
See exactly which external APIs, files, commands an agent touched
Bring Your Own Agent (BYOA)
Supported Agent Types
Agent Type
How It Connects
Claude Code
Direct adapter integration
OpenClaw
Full integration
Cursor / Codex
Adapter-based integration
Python script
Any script that can receive HTTP heartbeat
Shell command
Wrap in a heartbeat receiver
HTTP webhook
If it accepts POST, it’s an agent
Custom LLM
Any provider — OpenAI, Anthropic, Gemini, local
SKILLS.md — Runtime Context Injection
Agents use SKILLS.md to:
1. Understand what Paperclip tools are available
2. Learn project-specific conventions
3. Find where relevant context lives (codebase, docs, APIs)
4. Know how to delegate and escalate
# From zero to autonomous company in one commandnpx paperclip# Interactive setup guides you through:# 1. Database configuration (embedded Postgres or external)# 2. Auth setup# 3. Creating your first company# 4. Hiring your first agent# No Paperclip account required — fully self-hosted# MIT licensed — own every line of code
Deployment Options
Mode
Description
Best For
Local (embedded Postgres)
Single Node.js process, auto-manages DB
Development, solo use
Local (external Postgres)
Point at your own Postgres instance
More control, persistence
Remote/Cloud
Full cloud deploy
Production, teams
Self-hosted server
Docker or any VPS
Enterprise, privacy-sensitive
First Company Setup Flow
graph TD
Start["npx paperclip"]
DB["Configure Database"]
Auth["Setup Auth"]
Company["Create Company — Name + Mission"]
CEO["Hire CEO Agent — Choose LLM provider"]
Strategy["CEO proposes strategy → You review + approve"]
Team["CEO hires team → You approve each hire"]
Budget["Set budgets per agent"]
Go["Hit Go → Monitor dashboard"]
Start --> DB --> Auth --> Company --> CEO --> Strategy --> Team --> Budget --> Go
Pre-Built Company Templates
Template
What It Does
Content Business
SEO, blogs, social media on autopilot
Finance Firm
Analysis, risk assessment, compliance
eCommerce Store
Listings, support, inventory, ad optimization
YouTube Channel
Scripts, edits, thumbnails, scheduling
Software Dev Team
PM, engineers, QA, DevOps — full dev pipeline
Sales Agency
Prospecting, outreach, follow-up, closing
Problem → Solution
Before Paperclip
After Paperclip
20 Claude tabs open, can’t track which does what
Tasks are ticket-based, threaded, sessions persist across reboots
Manually gather context to remind your bot what you’re doing
Context flows from task → project → goal automatically
Folders of agent configs, reinventing task management each time
Org charts, ticketing, delegation, governance out of the box
Runaway loops waste hundreds of dollars
Token budgets surface + throttle agents when they hit limit