HOL Guard

You asked the coding assistant to tidy up one folder, and you would like to keep the rest of the disk. AI coding assistants can now run commands on your computer without asking, and one wrong step can delete a folder or send saved passwords to an outside server. HOL Guard sits between the assistant and the computer and checks each covered action before it runs. Normal work goes ahead, while risky actions are stopped or held until you confirm them.

The checks cover shell commands, reads of files that hold passwords or keys, package installs and calls to MCP servers, the add-ons that give an assistant extra tools. Guard also screens for text hidden in a page or file that tries to give the assistant new orders, known as prompt injection. Each decision is saved on your machine, so you can look back at what was blocked and why.

Guard is a command-line program that needs Python and is installed with pipx. A setup wizard finds the assistants already on the computer and asks before each change. It works with Codex, Claude Code, Cursor and more than ten others. How much it can stop depends on what each assistant reports before it acts, and the project publishes a coverage table for every one.

The local part is open source under the Apache-2.0 licence and keeps working offline without an account. The homepage says the local checks never upload your files. Guard Cloud is a separate, optional service that adds shared rules and decision history for teams.