HOL Guard
HOL Guard is a local runtime security layer for AI coding agents. It watches what agents attempt to run, read and install, asks for approval when needed and keeps a record of blocked and allowed actions. The free local component works directly on your computer without requiring an account or internet connection.
Coding agents like Claude Code, Codex and Cursor can access files, run shell commands and install packages. Guard intercepts supported actions before they execute, applying your policies to decide whether it runs, waits for your approval or gets blocked. Common protections stop deletion of files, access to credentials and API keys, exfiltration of data, installation of suspicious packages and risky shell commands like direct pushes to your main branch.
How Guard works
Guard Local runs on your computer and reviews supported tool calls at pre-action checkpoints. Your policy determines the decision path: allowed actions proceed immediately, ambiguous actions pause to ask for approval and blocked actions stop before execution. Guard keeps an audit trail of all decisions so you can review what your AI agent attempted to do and when. Decisions happen on your machine, and local protection keeps working offline.
Guard supports popular coding agents across macOS, Linux and Windows. Coverage varies by agent, since each has different integration points where Guard can intercept actions. A public coverage matrix shows what Guard blocks and observes for each agent, and what falls outside its scope.
Guard Free and Cloud plans
Guard Local is free and open source under the Apache-2.0 license. You can install it and read its code without a paid subscription or cloud account. Optional Guard Cloud adds shared policies and team dashboards, plus history across devices, on top of the local layer. A Solo plan, with a short free trial, keeps your decision history across devices, while Pro and Team plans add organization-wide policy enforcement and shared review dashboards.
Who needs Guard
Guard is most relevant if you let coding agents access your project files and run terminal commands or plugins. It provides a place to define what AI should and should not attempt on your machine. Because decisions happen locally before actions run, you can use Guard without routing your work through a third-party service. Developers working alone can use Guard Local for free, while teams managing policies across multiple machines benefit from the cloud coordination layer.
Security and privacy
Guard focuses on pre-execution control, not post-action recovery. It protects against risky actions your agent might attempt but does not guarantee every unsafe instruction will be caught. Keep your agent's own permission controls and regular backups in place alongside Guard. The security model addresses threats including secret exposure, destructive commands, prompt injection driven actions, malicious package installs and unsafe MCP configuration.
Alternatives to HOL Guard
- Agen: Helps teams discover and govern AI agents through a managed MCP gateway, with approval of sensitive actions and audit trails rather than local runtime interception.
- Orca: Runs terminal coding agents in isolated Git worktrees with built-in review and source control, providing process-level isolation instead of action-level policies.
- boat: Gives AI agents full Linux VMs with Docker, browser and desktop access, using sandboxed virtual machines for agent containment rather than local policy enforcement.
- Antigravity CLI: Brings Google's coding agents to the terminal with parallel subagents, plan and diff review and permission controls, emphasizing review workflows rather than pre-action blocking.
- iDox.ai Guardrail: Focuses on preventing sensitive information from reaching AI services through data detection and anonymization rather than blocking specific agent actions.

TRY IT