OpenCLAWN
An open-source control plane that sits on top of an AI agent's own logic, deciding whether each action is allowed, pausing it for human approval, and logging proof.
OpenCLAWN checks every tool call against policy before it can run, with a required-approval flag on risky actions that blocks destructive operations from executing before the check clears, rather than only logging them afterward. Every routing decision, tool call and skill promotion is logged both before it happens and finalized after, intended as a real audit trail rather than best-effort logging.
A routing-audit and self-calibration layer records why a routing decision was made and whether it was correct; a skill-decay mechanism keeps unused skills from accumulating and polluting context; confidence-gated "crystallization" is meant to stop the agent from storing skills learned out of bad solutions; and typed role-output contracts are used for multi-agent handoffs. On top of these, the skill library can tidy and improve itself as it is used, with the changes gated and reversible.
It supports multi-agent conversation patterns such as pipeline, debate and orchestrator roles with live stop-and-interject control, and splits one goal into an explicit, dependency-aware graph of subtasks.