Paperclip: Running an AI Agent Company in Claude Code

What is Paperclip

Paperclip, indexed from paperclipai/paperclip, is an open source Node.js server and React UI that orchestrates a team of AI agents to run a business, not just to finish a coding task. The project frames its own scope directly: if a tool like OpenClaw is one AI employee, Paperclip is the company built around it, the control plane that hires agents and tracks their work. Flowy indexes 22 components for it: 14 skills and 8 agents, filed under automation.

Flowy indexes this plugin from its public repository. It does not host, modify, or bundle the code. The install command lives on the listing page linked at the top of this guide.

How the control plane skills run a company

Five skills form the part of Paperclip you talk to directly. paperclip is the base skill for the control plane API itself: checking assignments, updating issue status, posting comments, delegating work, managing routines, and calling Paperclip endpoints. paperclip-board sits above that, built for chat rather than raw API calls, for a board member who needs onboarding, company or agent management, approvals, task monitoring, cost oversight, or work product review.

Two skills turn intent into assigned work. paperclip-converting-plans-to-tasks breaks a Paperclip plan into an executable issue graph, working out specialty fit, dependencies, blockers, and what can run in parallel. paperclip-create-agent handles hiring: it inspects adapter configuration options, compares existing agent configs, drafts a new agent prompt or config, and submits a hire request, with governance built in rather than a plain create call.

The fifth, paperclip-task-bridge, lets a separate assistant called Hermes create, comment on, update, and list Paperclip tasks using scoped Paperclip API credentials, for when work needs to start outside Paperclip's own chat surface.

What the LLM Wiki skills do

Six skills maintain what the project calls the LLM Wiki, a cited local company wiki meant to outlast any single chat session. wiki-maintainer is the entry point for using its tools day to day. wiki-ingest turns a captured raw source into durable source, entity, concept, synthesis, index, and log pages. paperclip-distill does the same job for Paperclip's own activity, turning a cursor window or backfill of task history into standups, project pages, decisions, and history, without dereferencing assets. wiki-query answers a question from wiki pages and raw sources with citations, names gaps plainly, and can offer to file new synthesis when the answer is worth keeping. wiki-lint audits for contradictions, orphans, weak sourcing, broken links, missing concept pages, and index or log drift, and only reports what it finds. index-refresh rebuilds the index from the actual page tree and reconciles missing or deleted entries, without editing page bodies itself.

What handles memory, design, and content

para-memory-files is a separate, file based memory system built on the PARA method: storing and recalling facts, daily notes, entity records, weekly syntheses, and user patterns across sessions, apart from the wiki. design-guide documents Paperclip's own UI design system, for creating or modifying UI components, adding pages to the frontend, styling elements, or understanding the design language the product follows. content-calendar looks outward instead: it plans a weekly editorial calendar by mapping company goals to publishable topics, owners, status, and verification notes.

What the 8 agents do

Four agents are hiring templates rather than agents you run yourself. coder, qa, securityengineer, and uxdesigner describe the role a new hire should fill (implementing and testing code, reproducing bugs and validating fixes, owning security posture, or producing UX specs and reviewing interface quality) for paperclip-create-agent to read, not for calling directly.

The other four are narrow, bounded sweeps. token-auditor scans the UI source for hardcoded visual values, duplicate components, and shadcn replacement candidates, and only reads, never touching component files. codemod-runner is its follow-up: once that audit exists, it writes and runs the codemod scripts that replace the hardcoded values with token references, doing mechanical refactors only. recent-agent-reflection runs a bounded review of recently active agents and produces coaching proposals only, never changing another agent's live instructions without an accepted task interaction. refresh-stale-summaries is paused by default and spends nothing until an operator enables its schedule or runs it by hand; when it runs, it only regenerates summary slots whose scope has changed, and never mutates issues, workspaces, or code.

Who actually reaches for this plugin

This is not a plugin for adding one skill to an existing coding session. Its 22 components assume a company-shaped structure: a board skill for oversight, a task bridge for outside tools, hiring templates, a cited wiki, and PARA memory so decisions survive between sessions, plus a pair of bounded agents that keep the UI's own source clean. It fits a team running Paperclip as its agent operations layer, not a developer wanting one extra skill.