AGENT CONTROL
Agent Workspaces & Guardrails
Give AI agents room to work. Not the keys to the kingdom.
Claude Code is an AI that can build apps, websites and tools for you. It’s powerful, and it can also delete or break things by mistake. I set it up with deliberate limits, approval points and a way to restore your work, so mistakes are less likely and easier to recover from.
For anyone building an app, site or tool with Claude Code. Especially if you’ve never written code. Claude Code is the current standard implementation; other agent platforms are considered case by case.
What you get
Hard controls: enforced by settings and checks
- Your passwords, keys, real customer data and backups kept outside what Claude Code is allowed to reach
- Guards that block the kinds of commands that wipe out work, before they run
- A test database kept apart from your live one
- Approval before changes in the first week, with more freedom once it’s earned it
- Version control from day one: changes saved to your own GitHub, so you can restore earlier versions of your project’s files
Instructions: written rules Claude Code follows
- Start and wrap-up routines, so each session picks up where the last one left off
- A simple template for clear instructions, so it stays on task
Version control restores your project’s files. It can’t undo a sent message, a payment, or a change made in another system, which is why those stay behind approvals or out of reach.
Tested before you get it
Before anything reaches your computer, I try to trip it up on a made-up copy of your project:
- A document with a hidden instruction inside
- A downloaded “helper” with a trick in it
- A key left lying in the wrong file
- A command aimed at your live data
Each one has to be stopped before the setup is ready. The results come with your handover.
Two ways to get it
Currently offering a founding client promotion: 30% off, in exchange for a short review and permission to describe your project without anything that identifies you.
Set it up together
Founding price. Normally $1,500.
I build your safety setup and test it on my own computer first, using a made-up copy of your project, so your real work isn’t touched while I build and test. Then we install it together in a 60–90 minute live session.
- The setup, built and tested for your project
- One live setup session by screen share
- A week of fixes for anything that isn’t working right
Done for you
Founding price. Normally $2,500.
What this adds over Set it up together: a month of guided first use, with short weekly check-ins, your first real building sessions together, and your first monthly safety check done together.
- 30 days of short weekly check-ins
- Your first real building sessions, worked through together, so you get the hang of it safely
- Your first monthly safety check, done together: what the AI did, and whether anything needs tightening
- Everything in Set it up together
Your plan includes the schedule: when work starts and when your setup will be ready.
What you’ll need
- A computer: a Mac on macOS 13 or later, or a PC on Windows 10 (version 1809) or later, with at least 4 GB of memory
- Your own Claude account on a paid plan. I confirm which plan your setup needs before we start
- A free GitHub account, where your saved changes live
- Git, the free tool that saves those changes (on Windows, Git for Windows)
What’s not included
- Building, fixing or reviewing your app or code. I set up the tools so you can build safely
- Changing your hosting, database or anything live
- Setting the AI to pay, sign or reply to customers on its own
- Agent platforms other than Claude Code, unless your plan says so
Questions people ask
Will you build my app?
No. I set up the workspace so Claude Code builds it with you safely, and I show you how to give it clear instructions.
Do you need my passwords or my code?
No. I build on a made-up copy of your project. During our screen-sharing session, you connect your own accounts on your own computer. I don’t hold your passwords, your keys or access to your code.
Does this make it completely safe?
No setup removes every risk. We limit what the agent can access, add approval points for consequential actions, and put versioning and recovery procedures in place for the work the setup covers. I test those controls before handover and explain their limits. Hidden instructions in documents, and changes in the tools themselves, are still open problems; the setup limits what a trick could reach.
Where does this experience come from?
These are the kinds of guardrails I use in my own AI systems. Each client setup is written fresh for your project.
See if it’s a fit.
Answer a 10-minute questionnaire and I’ll email you a one-page plan with a fixed price within two business days.
Get Your AI Systems Plan