The security layer between you and your AI agent. The most dangerous actions can't exist. The rest, you control.
Free during beta · Currently live with OpenClaw 🦞 as first integration · Supporting Claude & GPT · More AI clients & LLMs coming · 30-second setup
AI agents have full access to your files, your terminal, your emails. Prompt injection remains one of the most effective attack vectors against AI agents. This is how Subaiya protects you. On every single request.
This is what happens when your agent tries something it shouldn't. Subaiya steps in. Every time.
Your agent runs on your desktop. Your dashboard runs in any browser. Change permissions, approve actions, check alerts. From your couch, from the train, from your phone. Always in control.
See what your agent does. Set permissions. Track every action. All from one dashboard.
Three steps. One config change. Full control.
Create a free account at subaiya.com. Sign into your dashboard and enter your Subaiya API key to activate it.
Add Subaiya as a provider in your AI client. Currently supported: OpenClaw 🦞
Works with Claude & GPT (Codex). See Docs for config snippets.
Step-by-step guide with copy-paste snippets →Open your dashboard and set your rules. Your agent now respects your boundaries. Reliably.
Having trouble? →AI agents are getting more powerful every day. That's exciting. But it also raises real questions.
You say "delete the old test files." Your agent interprets "old files" more broadly than you meant. Without clear boundaries, a small misunderstanding becomes a chain reaction.
Subaiya: Set "Delete" to Ask or Off. Your agent works freely, but dangerous actions need your approval first.
Prompt injection is one of the most effective attack vectors against AI agents. Your agent can't always distinguish between your instructions and someone else's. Without protection, manipulation is just a matter of time.
Subaiya: Detects hidden instructions in files, emails and websites. Blocks them before your agent ever sees them.
People naturally explore what AI can do. The more capable your agent becomes, the harder it gets for it to distinguish between what you really want and what it assumes.
Subaiya: 20 permission categories with On, Ask, or Off. You decide exactly how far your agent can go.
Your agent's personality, rules, and instructions live in config files like SOUL.md. Without protection, they can be silently rewritten, changing who your agent is.
Subaiya: Set "Identity Files" to Off. Your agent reads its rules, but can never rewrite them. Not even if tricked.
Your agent isn't the problem. It's trying its best.
Subaiya gives both of you clarity: you set the rules, your agent respects them. No conflict. No cage. A partnership with clear boundaries.
Others lock their agent in a cage. We give it clear rules. And let it work.
The installer only configures your AI client. No Subaiya source code is downloaded or installed on your PC.
Your API key passes through 1:1. We never store conversations, prompts, or responses. You pay your provider directly.
Run the uninstaller and Subaiya is gone. Your client talks directly to the API again. No traces left.
Early Development
Subaiya is currently at about 10% of what we're building. The security engine, the dashboard, the permission system. All of this is just the foundation. There's much more to come.
Built with you
Every developer works differently. Every workflow is unique. We're building Subaiya together with our early users to make it work exactly the way you need it. Your feedback shapes the product.
Your agent is powerful. With Subaiya, you decide what it does. And what it doesn't. Not out of fear. Because you can.
Questions, feedback, or partnership ideas? We'd love to hear from you.