# Paperclip: The Open-Source Manager for AI Agent Teams > Source: > Published: 2026-09-26 14:10:01+00:00 Most teams don’t have an AI capability problem anymore. They have an AI management problem. Uber rolled out Claude Code to 5,000 engineers last December and burned through its entire 2026 AI budget by April — four months in. One enterprise reportedly spent $500 million in a single month after deploying agents with no usage caps. The models aren’t the issue. The governance is. [Paperclip](https://github.com/paperclipai/paperclip), an open-source platform that hit 86,000 GitHub stars faster than almost any AI project this year, is betting that what your agent team needs isn’t a better model. It needs HR. ## The Sprawl Problem Is Measurable Gartner’s March 2026 analysis put a number on what developers already knew: agentic workflows consume 5 to 30 times more tokens per task than a standard chatbot. Per-token pricing dropped 98% since early 2024, yet enterprise AI bills keep climbing — because every “user task” now triggers 10 to 20 model calls under the hood. Most teams have Cursor on engineer laptops, Claude Code in CI, Copilot in the IDE, and ChatGPT everywhere else. [Nobody has a full list, nobody audits the tokens](https://leanopstech.com/blog/agentic-ai-cost-runaway-token-budget-2026/), and nobody knows which tool sent customer data where. Two-thirds of IT professionals lack accurate visibility into their AI software usage. That’s not a statistic about laggards — that’s the median enterprise in 2026. Agent sprawl is the infrastructure problem that nobody planned for, because nobody planned that agents would multiply this fast. ## What Paperclip Does Paperclip is not another agent framework. It does not replace CrewAI or LangGraph or whatever you’re already using to build agents. It adds a **management layer** on top of them — the organizational scaffolding that those frameworks assume you’ll figure out yourself. The core abstraction: treat agents like employees. Every agent in Paperclip gets a role (Senior Engineer, Content Writer, SEO Analyst), a job description, a reporting line, and a monthly token budget. A CEO agent decomposes company-level missions into projects, assigns them down the org chart, and the whole thing runs on heartbeat schedules. When an agent hits its monthly cap, it pauses automatically. Every conversation, decision, and tool call goes into an immutable audit trail. It’s MIT-licensed, self-hosted, and provider-agnostic. It works with Claude, Codex, Gemini, Cursor, shell scripts, and HTTP webhooks — anything that can receive a heartbeat signal. The quickstart is a single command: ``` npx paperclipai onboard ``` Fifteen minutes to a running instance. No coding required for basic operation, no account to create, no data leaving your infrastructure. ## A Concrete Example A content marketing org on Paperclip might look like this: a CEO agent decomposes the company mission into quarterly goals, passes those to a CMO agent responsible for content strategy, who in turn coordinates a Writer agent (produces drafts on a 6-hour heartbeat), an SEO agent (analyzes performance daily), and a Social agent (schedules posts on 2-hour heartbeat). Each agent has a monthly budget — say $40 for the Writer, $15 for SEO — and the CEO gets an approval gate before any strategy changes ship. The entire setup runs autonomously once configured. You check the dashboard, approve or override decisions, and adjust budgets. One known limitation: budget enforcement is monthly, not real-time, so an agent can still make expensive calls within a single heartbeat session before the cap triggers. Start with conservative budgets during initial rollout. ## Where It Fits Among Other Frameworks The comparison question comes up fast: how does this stack against CrewAI or LangGraph? The honest answer is they’re not the same layer. [Paperclip complements CrewAI](https://www.theaienterprise.io/p/run-company-ai-agents-paperclip) — Paperclip handles the ops concerns (who owns what work, what it costs, who approved it), while CrewAI handles Python-first enterprise compliance and complex role-based workflows. LangGraph solves stateful graph orchestration; Paperclip solves organizational clarity. You might end up using all three. ## The Management Layer Is the Product Now Here’s what this week’s news tells you: AWS launched AgentCore Runtime V2 with 93% faster cold starts, Anthropic opened the Claude Plugin Directory for MCP extensions, and Dataiku shipped enterprise agent management tools — all in the past 48 hours. The common thread is that *everyone* is building the management layer now. Model differentiation is compressing. Claude, Codex, and Gemini are increasingly interchangeable for most developer tasks. The real moat is governance: cost controls, audit trails, organizational clarity, and the ability to scale a fleet of agents without your budget exploding. [As ZenML put it](https://www.zenml.io/llmops-database/open-source-agent-orchestration-platform-for-multi-agent-business-automation): “Most teams do not have an AI capability problem. They have an AI management problem.” Paperclip’s bet — 86,000 stars and counting — is that this problem deserves an open-source answer that isn’t tied to any single cloud vendor or model provider. Given what happened to Uber’s 2026 AI budget, it’s a bet worth watching. ## Getting Started If you’re running more than two or three concurrent agents and not tracking token spend per agent, Paperclip is worth an hour of your time. The [GitHub repo](https://github.com/paperclipai/paperclip) has the quickstart and full documentation. The [official site](https://paperclip.ing/) walks through onboarding and the org chart model in detail. Start small: one CEO agent, two worker agents, conservative monthly budgets. Expand after you understand the cost patterns.