Paperclip
Open-source orchestration platform for autonomous AI agent companies
Node.js server and React UI that orchestrates teams of AI agents to run businesses autonomously. Manages org charts, budgets, governance, goal alignment, and agent coordination through a task-manager interface.
Overview
Overview
Paperclip is an open-source control plane for autonomous AI agent companies. It provides the orchestration layer needed to coordinate multiple AI agents (Claude Code, OpenClaw, Codex, Cursor, etc.) toward common business goals. The platform handles org charts, budgets, governance, goal alignment, and agent coordination through a task-manager interface.
Architecture
The system consists of a Node.js Express REST API server and React UI. Core subsystems include:
- Identity & Access: Two deployment modes (trusted local or authenticated), board users, agent API keys, company memberships
- Org Chart & Agents: Hierarchical agent structures with roles, titles, reporting lines, permissions, and budgets
- Work & Task System: Ticket-based issues with atomic checkout, execution locks, blocker dependencies, and full audit trails
- Heartbeat Execution: DB-backed wakeup queue with budget checks, workspace resolution, secret injection, and skill loading
- Workspaces & Runtime: Project workspaces, isolated execution environments (git worktrees), and runtime services
- Governance: Board approval gates, config versioning, rollback capability, and agent termination controls
Agent Integration
Paperclip uses an adapter pattern to integrate any agent that can receive heartbeats. Built-in adapters support Claude Code, Codex, CLI agents (Cursor, Gemini, bash), and HTTP/webhook bots (OpenClaw). External adapters can be loaded as plugins via configuration files.
Cost Control
Monthly budgets are enforced per agent with automatic pause when limits are hit. All token usage is tracked and surfaced in the dashboard. Atomic execution semantics prevent runaway loops and double-work.
Multi-Company Support
A single deployment can run multiple companies with complete data isolation. Every entity is company-scoped with separate audit trails. Board users can manage portfolios of autonomous businesses from one control plane.
Development
The repository uses pnpm workspaces with packages for database schema (packages/db), shared types (packages/shared), adapters (packages/adapters), and plugins (packages/plugins). Development mode uses embedded PGlite for zero-config local setup. Full test suite includes Vitest unit tests, Playwright E2E tests, and release smoke tests.
Primitives
Orchestrates multi-agent teams toward business goals
Manages agent hierarchies and reporting structures
Enforces monthly budgets per agent
Tracks costs and token usage across agents
Schedules recurring agent work via heartbeats
Provides ticket-based task management
Maintains audit logs of all agent actions
Supports agent onboarding and configuration
Injects runtime skills and context
Handles workspace isolation and git worktrees
Outcomes
- 01Autonomous 24/7 business operations
- 02Controlled AI spending with hard budget limits
- 03Complete audit trail of agent decisions
- 04Coordinated multi-agent execution
- 05Persistent agent state across sessions
- 06Goal-aligned agent behavior
Integrations
Autonomy & guardrails
- Budget limits enforce automatic pause
- Board approval required for strategic changes
- Human governance via board interface
- Agent permissions scoped by role
- Company boundaries enforced at data layer
Guardrails & requirements
Guardrails
- Monthly budget hard-stops
- Approval gates for governed actions
- Atomic task checkout prevents double-work
- Board-level override and termination controls
- Company-scoped data isolation
- Activity logging for all mutations
- Rollback capability for config changes
Requirements
- Node.js >=20
- PostgreSQL or PGlite (embedded dev mode)
- pnpm 9.15.4
- Git for workspace management
Technical specifications
Runtime
- Harness
- Custom
- Deployment
- Self-hosted Node.js server. Two modes: trusted local (no auth) or authenticated (multi-user). Supports multi-company isolation in single deployment.
- Data residency
- Self-hosted with complete data control. Company-scoped data isolation. Secrets stored hashed at rest.
- License
- MIT
Models & tooling
- Tooling
- Express REST APIReact + Vite UIDrizzle ORMPGlite (embedded PostgreSQL)Git worktreesPlaywright for E2E testingVitest for unit testing
Security & compliance
- Auth model
- Board users have full-control operator context. Agent access via bearer API keys (hashed at rest). Short-lived run JWTs for execution. Company membership controls access boundaries.
Architecture notes
Memory architecture
Persistent task context and session state stored in PostgreSQL. Agents resume work across heartbeats with full goal ancestry and project context. Activity logs provide immutable audit trail.
Context strategy
Goal-aware execution where tasks carry full ancestry (company → project → goal → parent task). Runtime skill injection provides Paperclip workflows and project context without retraining. Workspace resolution ensures agents work in correct directories.
Evaluation
Vitest unit tests, Playwright E2E tests, release smoke tests, and promptfoo evals for agent behavior verification.