Select a principal
Identify the agent, extension, or delegated tool making the request.
For AI-agent security
Give autonomous tools less ambient authority.
Explore the permission model for files, commands, network destinations, secrets, tools, and delegation before the native runtime is built.
From extension identity to a recorded decision.
Describe the grant
Bind capability, action, resource, workspace, and expiration.
Audit the result
Record allow, deny, or prompt with a stable reason and policy version.
Less noise. More decision context.
Patch-first writes
The runtime target makes agent changes reviewable before applying them.
Brokered destinations
Network access is designed around explicit destinations rather than ambient egress.
Delegation visibility
Agents and tools keep distinct principals so authority cannot silently expand.
Start with a real workflow
Get started Explore the GuardRails permission prototype.
Use the working product surface now; no future runtime capability is presented as already shipped.