Git
Commands that rewrite or remove history.
Force pushHard resetBranch deletion
Keep AI agents away from dangerous commands and protected files
Alesia checks what an AI agent tries to run or open, and blocks dangerous commands, protected files, and anything your policies forbid.
An AI agent tries four actions. Alesia allows a test run and blocks a dangerous delete, reading a protected .env file, and a force-push forbidden by policy.
Why it matters
An agent can delete data in a single step.
Coding agents run commands and open files on their own. That is what makes them useful. It also means one wrong step can wipe a folder, rewrite history or put your secrets into a prompt.
Without Alesia
A coding agent is asked to clean up a build folder.
It runs rm -rf on the wrong path.
With Alesia
Alesia checks the command on the device before it runs.
The command is blocked. Nothing is deleted.
What it stops
Built-in detectors for destructive commands
Some commands are destructive no matter who runs them. Built-in detectors cover git, the file system, databases and cloud infrastructure.
Filesystem
Commands that remove files or change access.
Recursive deletesPermission changes on system paths
Databases
Statements that remove tables or their data.
DROPTRUNCATE
Cloud infrastructure
Commands that tear down running infrastructure.
terraform destroykubectl delete
Template
Start from Safe Coding Agent. It switches on the destructive git, filesystem and database checks. Or pick detectors one by one.
Files and your rules
Protected paths and your own commands
Agents read files to build context. Some files should never become context.
.envProtected by default
.env files are off limits from day one. The secrets inside never enter a prompt, a log or a commit.
~/.sshYour own protected paths
Add the paths that matter to you, such as your SSH keys.
./scripts/deploy-prod.shYour own commands
Add commands and patterns specific to your team. They are matched exactly.
How it works
Checked before the agent acts
Every command an agent tries to run and every file it tries to open is checked by Alesia on the device first. The check happens before anything executes, so a blocked action never touches your system. Actions that pass go ahead as normal, and the agent keeps working.
Dynamic rules
Describe risky behavior in plain language
Some risks don't fit a pattern. Describe them in a sentence, and the model checks agent actions against it. Test cases show whether a rule catches what it should and lets safe actions through.
Use cases
See it in action
Agent Guardrails, in detail
Agents that run commands or read files on a device with Alesia installed, such as coding assistants in the terminal or IDE and automation agents.
No. .env files are protected from day one, and the Safe Coding Agent template covers destructive git, filesystem and database commands. Add your own protected paths and commands on top.
Yes. Set the policy to watch, and matches are recorded as incidents while agents keep working. Switch to block once the rules look right.
No. Checks run on the device. Only the incident, such as which rule matched and what happened, reaches your console.
Say yes to AI.
Set the rules once. Alesia enforces them on every device, and prompts stay on your devices.