Paperclip

https://paperclip.ing/ · GitHub · Docs · MIT license, self-hosted

What it is

Paperclip is a self-hosted human control plane for fleets of AI agents. It models the runtime as a company: agents have roles, reporting lines, and job descriptions inside an org chart; goals cascade from company mission down through projects to individual tasks; tickets are the unit of work.

It is not an agent framework and does not prescribe how agents are built. Any agent that can receive a heartbeat can be plugged in. It sits above the harness layer and manages the organizational structure those harnesses operate inside.

Core capabilities:

• Org chart hierarchical agent structure with roles, bosses, and reporting lines.
• Goal alignment every task carries full goal ancestry so agents always see why they are working on something.
• Heartbeats agents wake on a schedule, check their work queue, and act. Delegation flows up and down the org chart automatically.
• Ticket system structured tasks with threaded conversations, full tool-call tracing, and an immutable append-only audit log.
• Cost control per-agent monthly budgets enforced atomically; agents stop when the limit is hit.
• Governance approval gates for agent hires and strategy changes; any agent can be paused, overridden, or terminated at any time.
• Multi-company one deployment can run isolated companies with separate data and audit trails.
• Persistent agent state agents resume the same task context across heartbeats instead of restarting cold.
• Runtime skill injection agents can learn workflows and project context at runtime without retraining.
• Portable templates org structures, agent configs, and skills can be exported and imported.

Why it fits Ikidna

Paperclip directly addresses several Minimal Bootstrap requirements in one layer:

• Heartbeat scheduling provides the cron-driven re-invocation mechanism.
• Goal ancestry satisfies durable, cross-invocation goal definition.
• Ticket system functions as issue tracking for long-term task decomposition.
• Budget enforcement is a concrete cost-control and continued-existence safeguard.
• Governance gates support the controls and approval model.
• Full tool-call audit log feeds the execution ledger.

It also aligns with the bring-your-own-agent model Ikidna requires no agent lock-in, any harness that accepts a heartbeat can be hired into the org chart.

Cons

• With most non mature AI software solutions, it focuses on ease of use and local first execution and not the level of scaling and segregation that it needs.

  • The full system should not need to be redeployed to add new tools, etc.
  • The system should be broken up to support better scaling and boundary segregation

• The feedback loop can result in cascading failures and token exhaustion by continually retrying the system without an escape hatch

• Cost control does not integrate well with external systems like an AI gateway