HOL Guard is a runtime firewall for AI coding agents, and the word runtime is the point: it sits between the agent and the machine, inspects each tool call as it is attempted, and stops the dangerous ones before they execute. The actions it watches are the ones that actually cause harm in an agent session, namely deleting files, reading credentials, installing packages and sending data out. Known-bad calls are blocked outright, ambiguous ones pause for your approval, and every interception is recorded as a security receipt. It runs locally, without a cloud account and without an internet connection, so nothing on the machine is uploaded in order to be checked, and it hooks into the harnesses developers actually use: Claude Code, Codex, Cursor, OpenCode, Gemini CLI and Hermes. The Guard Local core is open source under Apache-2.0, which matters more here than in most categories, because a security layer you cannot inspect is a security layer you are trusting blind. Pricing starts at a free forever tier, with Solo at $4.99 a month, Pro at $15 and Team at $30 a seat. The paid tiers buy Guard Cloud: synchronised policy, fleet visibility and team workflows, none of which the local runtime needs to do its job on one machine.
Zenity
4.2Security and governance for AI agents, from build time through to runtime.