Paperclip: Open Source Human Control Plane for AI Labor — Dotta Bippa

April 15, 202624:3446,609 views

Summary

Dotta (anonymous creator, known only by avatar) presents Paperclip, an open-source human control plane for AI labor, walking through the tool's architecture, capabilities, and vision for zero-human companies. Paperclip is positioned not as a fully autonomous business-creation tool, but as a system where humans remain in control of the higher-level design and execution of agent-driven work — a distinction Dotta emphasizes repeatedly throughout the talk.

The core metaphor is an org chart of AI agents. Users define a company with a CEO agent, who in turn delegates to an executive layer (CTO, CMO) and individual contributors (coders, content strategists, video writers). A key technical differentiator is the bring-your-own-agent model: Paperclip is vendor-neutral, supporting Claude Code, OpenAI Codex, Gemini, Hermes, Pi, Cursor, and models via OpenRouter — including free-tier models like Qwen 3.6+ for lower-stakes tasks, reserving frontier-model spend for high-intelligence work.

The talk demonstrates a live Paperclip instance used to manage the Paperclip project itself. Dotta walks through a real example: he created a task asking the CEO to hire a video writer and produce a Remotion-based animation celebrating 40,000 GitHub stars (the project crossed 50,000 stars live during the recording, approximately 34 days after the March 4th launch). The video writer agent was automatically equipped with a Remotion best-practices skill and access to the Paperclip branding guide and stats dashboard — context that would normally require manual assembly if using Claude Code directly. The resulting animation was produced in roughly 5 minutes.

Skill management is a central concept. Skills are reusable instruction packages — Paperclip integrates with skills.sh and supports both generic skills (e.g., Remotion best practices, the agent browser skill for web automation) and organization-specific skills (e.g., Paperclip branding preferences, pacing choices). A skill consultant meta-agent is used to diagnose when agents underuse their assigned skills and iterate on them.

Reliable workflows are handled through configurable task review and approval chains. When a coding agent finishes a task, a QA agent with the browser automation skill must review before the manager approves — enforcing quality gates without relying on individual model compliance. This is contrasted with the fragility of hooks in Claude Code or Codex that work differently per tool. Routines add scheduled or template-driven automation: daily changelogs, Discord merge summaries, PR reviews via the Greptile skill.

On cost management, Paperclip tracks monthly spend per agent and per project, with budget caps. At the time of recording, Claude and Codex subscriptions are used instead of API billing, but Dotta notes this changes at scale.

The roadmap announced includes: CEO chat, maximizer mode (continuous autonomous execution when tokens and ambition permit), multi-human workspace support (being built that week), sandboxed agent execution via E2B or Daytona, a desktop app, cloud deployments, improved memory and knowledge base handling, and artifact/deployment management. Dotta frames all of these as coming within 30 days of the recording. The project is installable today via `npx paperclip-ai onboard` and is fully open-source at paperclip.ing.

Key Takeaways

•Install Paperclip with `npx paperclip-ai onboard` — no technical background required to manage AI agent employees.
•Model Paperclip as an org chart: CEO delegates to executives, who delegate to individual contributors; agents negotiate, communicate, and share memory within the system.
•Use vendor-neutral agent routing — assign cheap or free models (e.g., Qwen 3.6+ via OpenRouter) to lower-stakes tasks and reserve frontier models (Claude, Codex) for high-intelligence work.
•Build skills as reusable instruction packages specific to your organization (branding guides, pacing preferences) so agents produce on-brand output without manual context assembly.
•Enforce reliability through configurable review/approval workflows rather than trusting individual agents to self-police — a QA agent must sign off before a manager approves.
•Use routines for recurring agent tasks (daily changelogs, PR reviews, Discord summaries) with template variables instead of copy-pasting prompts.
•Start your Paperclip organization agent-by-agent; resist importing large 130-agent templates until you have crafted clear instructions for each role.
•Add a skill consultant meta-agent to diagnose when agents underuse their assigned skills and iteratively improve those skills over time.
•Set per-agent and per-project budget caps to control frontier model costs as your agent team scales beyond subscription limits.
•Paperclip is not a coding tool — it is designed for running any business function (marketing, sales, finance, operations) and is accessible to non-technical users.

Transcript

Welcome to getting started with Paperclip and I'm Doda, your host and the creator of Paperclip and I'm so excited to show you how to get started with this brand new open-source agent orchestrator. I'm going to walk you through your first steps of how to get Paperclip started, what it looks like when you have a huge organization, and some of my best advanced tips for working with AI agents and so you can create your own uh zero-human company. Now, the tagline that we have for Paperclip is that it...