Paperclip Review

Paperclip is an orchestration layer rather than an agent. It models a company: agents hold roles such as CEO, CTO, or engineer, tasks trace from the company mission through project and agent goals (with context in a SKILL.md file), and agents wake on heartbeats, for assigned work, follow-up messages, or schedules, check their queue, and act. The README lists adapters for OpenClaw, Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, Hermes, Grok Build, and Kimi Code, plus custom processes, HTTP endpoints, and external adapter packages. Budgets can be set per company, agent, or project, with warning thresholds and hard stops. Governance includes board approval workflows, execution policies with review and approval stages, revisioned configuration changes that can be rolled back, and an append-only audit log, and one deployment can run several isolated companies.

The quick install is npx paperclipai@latest onboard --yes, which needs Node.js 24.11 or newer. From source you clone the repository, run pnpm install and pnpm dev (needs pnpm 9.15 or newer), and the UI and API come up on localhost:3100. Signing in with a local Claude or Codex subscription also needs Python 3 and the provider CLI, and source builds may need a Rust toolchain unless you set PAPERCLIP_RUNNER_BINARY. Local runs use an embedded PostgreSQL instance created automatically, and for production you point Paperclip at your own Postgres. For remote access the README suggests authenticated mode bound to a private network such as Tailscale, or Docker. Private npm registries can cause E404 errors on npx, and the README gives a workaround.

You act as the board: approve hires, set budgets, review strategy, and pause or reassign agents, while the ticket system records every instruction and response. The README notes that a mention alone does not assign work or wake another agent, which keeps agent-to-agent chatter from spiraling. This section reflects the README; we did not run Paperclip.

• Org charts with roles, goals that trace to the company mission, and delegation through the hierarchy
• Heartbeat scheduling and routines that create recurring tasks
• Budgets per company, agent, or project with warning thresholds and hard stops
• Board approvals, execution policies, revisioned configuration, and an append-only audit log
• Adapters for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, Hermes, OpenClaw, and custom processes or HTTP endpoints
• Multi-company isolation in one deployment, plus plugins, a secrets manager, and an MCP tool gateway

• Spend control and governance are built in, which most agent tools leave to the user
• Runtime-agnostic: it coordinates other vendors' agents instead of replacing them
• Explosive growth: about 100,000 stars, up 25% in a month
• A clear roadmap in the README with shipped, in-progress, and not-started items

• Budget enforcement relies on recorded spend, and usage reporting or in-flight work can delay a stop, so spend can overshoot a limit
• Agent Chat and the chat and email connectors are experimental and must be enabled in instance settings
• Memory and knowledge, cloud deployments, a desktop app, bring-your-own ticket system (Asana, Linear, Jira), and work queues are not built yet
• Needs Node.js 24.11 or newer and pnpm 9.15 or newer, with optional extra tooling for source builds
• The README has no known-limitations section, and quality ultimately depends on the agents you plug in

Paperclip is free under the MIT license (copyright Paperclip Labs, Inc.) and requires no account. You pay for the model providers or agent services your agents use and for any server you host it on. The OpenAlternative page carries an affiliate VPS offer, which is not Paperclip pricing.

Claude Managed Agents is Anthropic's hosted infrastructure for running Claude agents, and Relevance AI and Lindy are commercial agent platforms. Paperclip is a self-hosted layer above agent runtimes: it does not supply models, it coordinates agents from many vendors and adds budgets, approvals, and audit trails. That makes it complementary to the agent tools it manages and an alternative to hosted agent orchestration platforms.

• Teams experimenting with multi-agent organizations and wanting budget limits
• Developers who run several agent runtimes and want one control plane
• Anyone who needs an audit log and approval gates around autonomous agents
• Builders who want an open base for agent-run companies

• Teams that want a single simple assistant
• Organizations that need a mature, vendor-supported platform
• Users who do not want autonomous agents spending money unattended