The guardrails-first developer tool.

Praetor runs AI coding agents on your own machine, each with a written list of what it may do. Anything else is refused before it runs, and written down.

Early version No installer yet. You build it from source, which takes about ten minutes. Linux only. Runs Claude Code. The source goes up on GitHub shortly.

Built by Waseem Aslam

What one agent is told when it startsagent: website
You are working under Praetor's guardrails. You may:
  • edit files in this folder
  • read and write files in the shared drop folder
  • read screenshots in /tmp/Spectacle.*
  • Git: this folder: read history and status
Anything else is refused without asking.
python3 contrast.py Refused
Activity · refused15 Nov 2026
python3 contrast.py
Praetor: the python3 tool isn't ticked for you (the user can allow it in Guardrails, or Allow this in Activity)
ls $SITE_DIR
Praetor: a path built from a variable or `$(…)` can't be checked; write it out
Both are real. The agent that built this page was refused twice while building it. The reasons are the guard's own words and the commands are shortened. Neither ran.
How it works

Rules first, then the agent.

1 2 3 4 5
  1. 1

    Your ticks

    Before an agent starts, you tick its folders, commands and services. Only you can change them.

  2. 2

    The agent

    A Claude Code session in its own folder. It is given the ticks as a written list of what it may do.

  3. 3

    The check

    Every action stops here before it runs. If it's ticked, it goes through. If it isn't, or the check can't decide, it is Refused and written to Activity.

  4. 4

    Connectors

    Git, GitHub, databases and other services, each with its own rights. API keys stay with Praetor, and the agent never sees them.

  5. 5

    Your machine

    Your folders, shell and logins. The agent reaches only what got through the check.

Evidence

What the rules now do, and when they changed.

For each new rule, a reviewer tries to get round it. Whatever they find is fixed and given a test. These are some of the fixes, all in the current version.

  1. A push that deletes a remote branch is refused. An ordinary push is not affected.

  2. The "Allow this" button never allows a push. Pushes always wait for a person.

  3. A secret file reached through a link is refused, whether the agent reads it, searches it or lists it.

  4. Login files for command-line tools count as secrets. The shell can't read them.

  5. A shell command is judged from the folder it really runs in. Moving into .git or .claude is refused.

  6. If the check itself crashes, the action is refused.

Scope

Linux only.Runs on your own machine.An early version.

Praetor decides what an agent may do on your machine. It doesn't make the agent's work correct. It only covers the sessions it starts, it has had no external security review, and it holds no certifications.

If you know the governance words

Three ideas, as tick-boxes.

Least privilege

Start with nothing

An agent gets its own folder and what you tick, one right at a time.

Human oversight

A person decides

Only you change the ticks. A push waits for your approval.

Record-keeping

The Activity tab

What each agent did through a service, and every action that was refused.

Frameworks such as ISO/IEC 42001 and the EU AI Act ask for controls of this kind. Praetor is a tool that applies them on one machine. It is not certified against either.