Paperclip: A Task Manager for Teams of AI Agents, Not Just Terminals

The Problem: 20 Terminals and No Oversight

The README paints a familiar scene: you have twenty Claude Code terminals open and cannot track which one does what. On reboot, you lose everything. Paperclip targets that chaos. It is for people who run many AI agents, from different providers, toward a common business goal. The pitch is direct: if OpenClaw is an employee, Paperclip is the company. That means tasks become tickets, conversations become threads, and sessions persist across reboots. The core problem it solves is not agent execution, it is coordination and accountability. It is for managers who want to approve work, set budgets, and audit what agents did, all from a dashboard. The README calls this 'managing business goals, not pull requests.'

How It Works: Heartbeats, Tickets, and an Org Chart

The architecture is a Node.js server with a React UI. Agents connect by receiving a heartbeat. The README says: 'If it can receive a heartbeat, it's hired.' This is the key mechanism. Agents wake on a schedule, check work, and act. Delegation flows up and down the org chart. Each task is a ticket, and every conversation is traced. The system records tool calls and maintains an immutable audit log. The four pillars structure the design: an agentic task manager, an org chart for agents, agent employee training, and an agentic OS. The task manager handles tasks, approvals, and review gates. The org chart defines roles, permissions, and scoped secrets. Training covers skills and evals. The OS provides cross-provider runtime, sandboxing, and cost controls. In practice, you define a goal, hire a team of agents, approve the strategy, set budgets, and monitor. The dashboard is the control plane.

Getting It Running: Quickstart and Configuration

The README points to a Quickstart section, though the full commands are truncated in the README. It links to docs.paperclip.ing for details. The project is a Node.js server, so you would clone the repository, install dependencies with npm, and start the server. The UI is React-based, so you likely run a build step. Configuration involves defining an org chart and assigning roles. The README mentions monthly budgets per agent. When an agent hits its limit, it stops. That implies a config key for budget per agent, though the exact key name is not given in the README. The release versions follow a date scheme, like v2026.824.1, suggesting a continuous delivery cadence. The repository has a master branch and is not archived. The license is MIT, so you can modify and redistribute freely. The homepage is paperclip.ing.

Cost Control and Governance: The Real Selling Points

Two features stand out: cost control and governance. Monthly budgets per agent are enforced. When the limit is reached, the agent stops. That is a concrete, hard stop, not a soft warning. This addresses the fear of runaway AI costs. Governance is equally direct: you can approve hires, override strategy, pause or terminate any agent at any time. The README lists this under the governance feature. This is not just a monitoring tool; it is a management tool with authority. The ticket system provides an immutable audit log, so every decision can be explained. For organizations that need to show compliance, this is valuable. The multi-organization support means one deployment can serve many orgs with complete data isolation. That is useful for agencies or portfolio companies managing multiple autonomous businesses.

Where It Might Be the Wrong Tool

The biggest limitation is the heartbeat model. If your agent cannot receive a heartbeat, it cannot be hired. That excludes agents that are purely event-driven or run in closed environments. The README lists supported agents: OpenClaw, Claude Code, Codex, Cursor, Bash, and HTTP. If you use a proprietary agent that does not expose a heartbeat endpoint, you are out of luck. Another limitation is the complexity of the org chart. If you only run two agents, setting up roles, reporting lines, and budgets is overkill. The README's own 'right for you if' list includes having 20 simultaneous terminals, which is a high bar. The training pillar, with evals and performance reviews, suggests a serious investment in agent development. If you just want to run a single script, this is not the tool. The README also mentions 'scoped secrets' and 'company boundaries,' which imply a learning curve for security configuration.

Alternatives: OpenClaw and Plain Orchestration Frameworks

The README explicitly frames OpenClaw as an employee, not a company. So OpenClaw is a direct alternative for the agent runtime itself. OpenClaw likely focuses on giving a single agent a persistent memory and tool use, while Paperclip focuses on coordinating many agents. The difference in approach is that OpenClaw is the worker, Paperclip is the manager. Another alternative is to roll your own orchestration with a task queue and a database. That is what Paperclip replaces, but it gives you less control. The README's comparison table shows the pain of manual orchestration: losing context on reboot, no audit trail, no cost limits. Paperclip's approach is to centralize all that. If you already use OpenClaw and just need a dashboard, Paperclip can sit on top. If you want a lightweight scheduler, you might prefer a generic workflow engine, but you would have to build the governance and budget features yourself.

Maintenance and Upgrade Cost

The project has recent releases, with v2026.824.1 pushed on 2026-08-25. That indicates active maintenance. The date-based versioning suggests a release cadence tied to the calendar, which could mean frequent updates. Upgrading is a matter of pulling the latest release and restarting the server. The MIT license means no licensing fees, but you must include the copyright notice. The README does not mention a migration path between versions, so you should check the docs before upgrading. The cost controls are per-agent budgets, but there is no mention of infrastructure costs, so you still pay for the Node.js server and any databases. The audit log is immutable, which means it grows over time; you will need to plan for storage. The training and evals features imply a data store for test runs, which also adds storage cost. The README mentions 'data privacy' and 'internal trace collection,' so you should review how traces are stored and whether they contain sensitive information.

Editorial conclusion

Adopt Paperclip if you coordinate many different agents (OpenClaw, Codex, Claude, Cursor) toward shared goals and need budgets, approval gates, and an audit trail. Skip it if you only run a single agent or need deep provider-specific features. Before adopting, verify that its heartbeat model fits your agents' runtimes and that your workloads tolerate the overhead of a task manager layer. Check the MIT license and the project's recent release cadence (v2026.824.1) to judge maintenance.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes