HOL Guard

An AI coding agent may need to run a build and delete files in the same session. HOL Guard lets a developer apply different rules to those actions where the agent exposes a supported hook. Its documented choices include allowing an action, blocking it or asking for approval, with a record of the decision afterward.

That makes it a candidate for developers who want repeatable action rules across supported coding tools. The useful distinction is between a policy decision made before execution and a log that only records what already happened. Guard's coverage documentation separates those capabilities.

The public coverage matrix lists event types, approval behavior and blind spots by integration. For example, its Codex entry says inline edits made without a tool call are invisible to Guard. Its OpenCode entry says file reads and writes bypass Guard unless OpenCode permission rules block them. A supported-agent label therefore doesn't mean every action is intercepted.

The site's headline promises are broader than those limits. The matrix is the better starting point for deciding whether Guard covers a particular workflow. No agent actions were run through Guard for this note, so its enforcement and resistance to malicious instructions remain untested here.

The pricing page says Guard Local is free and open source, with local approvals and decision records available without an account or internet connection. Guard Cloud adds paid synchronization and team features. That separation lets a developer evaluate the local integration before considering shared history or organization policy.

Teams considering Cloud should also read its data-handling terms. The pricing page describes syncing redacted decision summaries, while the local runtime evaluates actions on the developer's machine.