laya-guard is a local check that occurs before any coding agent tool call, deciding whether to approve, send for review, or block the call.
When Jev was released, I was curious as to how I could use it, and I quickly discovered Laya which is an open-source 322M model that functions almost identically but runs on a laptop's CPU in under 100ms, as the check is local, there won't be a chance of 3rd parties seeing your commands, directories, or tokens, and there's no charge per call, I often work in Korean, and most of the guardrails I've seen have not evaluated Korean at all, hence why I've implemented it as well.
Regex policies catch known dangerous patterns before even reaching models, any command that is blocked is never seen by models, commands like ls, git status, and npm test will skip the model altogether, unless they have pipes, chaining, redirects, -exec flags, or sensitive paths, everything else is sent to the model for processing, once you have the model downloaded, it makes decisions from a fixed set of outcomes with no API key or request needed for any external service, there are also hooks for 11 different agents including Claude Code, Codex, Cursor, and many more.
And Also a local dashboard where you can see all decision and why it was made and install the hook for each agent, it doesn't save the commands themeslves.
In my tests with AgentDojo, it blocked 25/27 attacks, and allowed 92/97 harmless tasks, it correctly identified 64/66 commands I made.
After the Reddit post, people have found some issues with it:
Adding a comment at the end of a command would fool the model to not block the command.
The model only read the first 4,000 characters, so it could be spammed with junk to the beginning to bypass the check, this would send dangerous commands like curl to review.
"curl" repeated 40,000 times caused one regex to run for 48 seconds, timing out the hook and running the command.
All 3 issues have already been fixed, somebody also pointed out that if the server was down, commands ran without any check, now it asks you first, or blocks the command if the agent can't ask
Please note that this is not a tool to check if the code written is correct, it's a tool to prevent agents from running dangerous commands.