HOL Guard sits in the newest and most necessary tooling gap there is: between an AI agent deciding to do something and the something happening. It reviews risky tool calls — file deletions, outbound requests, spend — before they run, locally, as free open-source software. Anyone who has watched an agent cheerfully rm the wrong directory understands the product in one sentence.
What it costs
Free and open source, local-first — the only architecture that makes sense for the job, since a cloud security gate for your local agent is a contradiction with latency.
What we can’t tell you
Coverage and friction, the eternal security trade. A guard that prompts too often gets clicked through like a cookie banner; one tuned too loose misses the one call that mattered. Check which agent frameworks it hooks into today (the integrations list is young) and expect to tune the risk rules to your own workflow — shipped defaults are a starting point, not a policy.
The details
| What it is | Local, open-source review gate for risky AI-agent tool calls |
| Price | Free, open source |
| Best for | Anyone running coding agents with real permissions |
| Think twice if | You’ll click through every prompt — an ignored guard is theatre |
