Paperclip Claude skills

Setting up a Paperclip company, step by step

Paperclip is a Node.js server with a React UI, and the README frames the workflow in three steps: define the goal, hire the team, then approve and run. Its own example goal is "Build the #1 AI note-taking app to $1M MRR," staffed with a CEO, CTO, engineers, designers and marketers, where each role can be any bot from any provider. The tagline "If OpenClaw is an employee, Paperclip is the company" sums up the model. You manage business goals and reporting lines rather than individual pull requests.

A typical first run looks like this:

1. Try a disposable instance. Set a provider key and run ANTHROPIC_API_KEY=... npx paperclipai test-drive. For Codex, use OPENAI_API_KEY=... npx paperclipai test-drive --harness codex. The README also shows a test-drive variant that takes OPENROUTER_API_KEY. The test-drive guide at docs.paperclip.ing covers how credentials are handled and reused.
2. Onboard a real instance. With Node.js 24.11 or newer, run npx paperclipai@latest onboard --yes.
3. Hire agents. Choose a model and harness for each agent: Claude Code, Codex, Cursor (including Cursor Cloud), Gemini CLI, OpenCode, Pi, Hermes (including Hermes Gateway) or OpenClaw. Tasks, skills, permissions and history for the whole team live in one place, so each role in the org chart can run on a different runtime.
4. Review strategy, set budgets, hit go. You approve the plan, attach budgets, and then follow work and spend from the dashboard.

A few setup problems are worth knowing about in advance:

• Signing in with a local Claude or Codex subscription, instead of an API key, also requires Python 3.
• If npx paperclipai fails with an E404 and you use a private npm registry such as GitHub Packages through a global ~/.npmrc, see the README's troubleshooting note on private registries. The global config is the likely cause.
• Telemetry is on until you turn it off. If you need that, set PAPERCLIP_TELEMETRY_DISABLED=1 before the first run so no events are sent.
• A hosted option is not available yet. The README links to a Paperclip Cloud waitlist, which means self-hosting is the only way to run it today.

How budgets, approvals and access control work

The README's main pitch is about cost. It describes the problem directly: runaway loops can burn hundreds of dollars in tokens and hit your quota before you notice. Paperclip responds with spend tracking, budget alerts and automatic pauses.

Cost tracking and budgets. Paperclip records tokens and cost along several dimensions: company, agent, project, goal, issue, provider and model. Budget policies are scoped, so you can put a warning threshold and a hard stop on one agent, one project or a whole company. Enforcement works from recorded spend. If you run more than one company on an instance, each one gets its own accounting.

Governance. Board approval workflows sit in front of strategy. Execution policies can add review and approval stages to work, and configured task reviews decide when a task counts as complete. Paperclip keeps a record of decisions, and you can pause, resume or terminate any agent. Atomic task checkout keeps two agents from picking up the same ticket.

Identity and access. There are two deployment modes: trusted local, or authenticated. In authenticated mode you get human roles and permissions, company memberships and invite flows. Agents receive their own API keys, and each run uses a short-lived JWT, so a run token expires shortly after the run instead of sitting around as a long-lived secret.

Heartbeats and continuous runs. Agents wake on heartbeats and keep working between check-ins. That is why the app fits unattended runs, but the budget hard stops matter most in exactly that case. Set them before leaving anything running overnight.

Tracing. Separately from usage telemetry, the server ships opt-in OpenTelemetry auto-instrumentation (traces only). It turns on when OTEL_EXPORTER_OTLP_ENDPOINT is set and supports the grpc, http/protobuf and http/json protocols. This is useful if you already run a tracing backend and want to see agent runs alongside your other services.

What the bundled maintainer skills actually do

Most skills under .agents/skills assume a Paperclip checkout. They usually find the repo through PAPERCLIP_ROOT or git rev-parse --show-toplevel, then confirm that expected folders such as tests/runner-e2e and packages/paperclip-runner exist. Here is what a sample of them covers:

• check-pr is the most portable skill in the repo. Its metadata credits greptileai (version 1.3). It inspects a GitHub PR, GitLab MR or Perforce changelist for review comments, failing checks and gaps in the description, then helps fix them. It needs gh, glab or p4 installed and authenticated. It detects the platform on its own, and you can pass --vcs gitlab for self-hosted GitLab hosts whose names don't contain "gitlab."
• add-product-e2e-eval adds full-stack workflows that use real Chromium, the Paperclip server, a database, the runner, a provider, and optionally Daytona. Workflows are registered in tests/runner-e2e/catalog.ts, and you confirm discovery with pnpm test:e2e:runner -- --list --suite <suite-name>. The skill tells the agent to keep secrets out of fixtures and screenshots, to use immutable digest references for Daytona images, and to make sure missing evidence never produces a pass.
• add-runner-eval covers protocol evals against a seeded mock control plane. Case definitions live in a sibling paperclip-evals repository, and you check them with python3 validators such as eval_program.py validate.
• verify-paperclip launches a local instance on a free port between 3100 and 4099 using a separate data directory. Before stopping a server, it checks which process owns it.
• garden-inbox archives inbox issues through the app API and prints an undo call for each archived item.
• company-creator works from a companies spec that links agents, skills and projects through AGENTS.md, SKILL.md and PROJECT.md files.
• create-agent-adapter, create-runner-adapter and create-external-bot-integration guide adding new runtimes and bots. The agent adapter guide alone is about 32 KB.
• create-paperclip-bundled-skill and create-verification-skill explain how to write new skills in the house format, with templates and reference examples.
• deal-with-security-advisory, diagnose-why-work-stopped, doc-maintenance and fix-slow-queries cover maintenance work. For slow queries, the skill tells the agent to try query and caller changes and existing indexes before proposing a migration.
• directions-lab is the largest skill by file count, with dozens of scripts, tests and templates.

If you want to borrow one, check-pr is the realistic choice. Treat the rest as worked examples of how a busy repo writes agent procedures: explicit repo discovery, required reading before edits, and narrow validation commands.