agent safety

DashClaw is a policy and approval layer for unattended AI coding agents. It intercepts risky or expensive actions before execution, routes them for remote approval, and records signed evidence for later review.

unYOLO is a framework for building credential brokers and proxies for AI agents. It keeps provider credentials out of the agent, applies local policy rules, supports timed grants and operator approvals, and includes brokers for GitHub, Hugging Face, and approved Unix commands.

Kernel-level read gate for AI coding agents that blocks protected file reads before bytes reach the agent, with Linux fanotify and macOS Endpoint Security support.

A command guard for AI coding agents that blocks destructive git and shell commands before they run, with support for multiple agent and CLI tools.

LeanGuard is a Lean 4-based runtime guardrail system for agent tool use, with native temporal policies, Python integration, and optional MCP support.

WorldCut is a TypeScript package that checks whether observations from independent systems satisfy declared version and time relationships before an autonomous agent makes a decision.

Open-source runtime action grounding framework for AI agents. It verifies proposed agent actions against live systems of record before execution and can return allow, hold-for-human, or block decisions.

A safety-evaluation harness for AI agents operating physical equipment. It validates proposed robot intent against declared hardware and deployment limits, records machine-readable refusal reasons, and includes a sample corpus plus a runnable evaluation script.