All tools / HOL Guard

Basic tool profile

HOL Guard: Local Policy Checks for AI Agents

HOL Guard adds a local policy checkpoint to supported AI-agent actions. It can allow an action, ask for approval or block it under the configured policy. Its usefulness depends on the agent integration and the specific events it can intercept.

Visit HOL Guard

Information is based on public sources, not hands-on testing; confirm current features and pricing with the vendor.

Check coverage before installing

The product documentation describes checks around supported shell commands, secret access, MCP changes and software or skill installation. A supported agent name does not mean every action is covered. Read the relevant integration guide and confirm whether each event is checked before execution or only recorded afterward. Its documented integrations include OpenClaw as well as coding-focused tools; check coverage for the agent and events you intend to use.

For example, the Codex integration guide describes command review and provides integration-test and diagnostic commands. The public repository is the source-code entry point.

Local use and optional cloud features

As of November 6, 2026, the pricing page separates free local use on one machine from optional paid Guard Cloud plans. Local decisions and approval history do not require a cloud account. Cloud plans add capabilities such as synchronized history and team policy; compare device limits, retention and current billing terms before subscribing.

The publisher says local evaluation happens on the machine and optional cloud sync uses redacted decision summaries. Review the settings and data policy for the mode you intend to use.

A bounded evaluation

  1. List the actions you need controlled and match them to the documented integration coverage.
  2. Start in a disposable workspace with harmless commands and check whether expected approval prompts appear.
  3. Compare the recorded decision with the action that actually ran. If coverage is unclear, pause the evaluation and inspect the integration diagnostics.

Guard is not a complete prompt-injection defense or a replacement for your agent's native approvals, sandbox and operating-system permissions. For software-development workflows, start with our coding assistant guide if you are still choosing the underlying agent.