astrid

Why we included this project

Astrid puts the security boundary for agent workloads in the runtime rather than in the model. Every component runs inside a sealed WebAssembly capsule, and the permissions it receives are signed, scoped capabilities that can be revoked. A jailbroken prompt, poisoned tool, or plain bug still cannot act beyond that grant, whether that means touching a file or opening a network connection. The kernel is small by design, routing events and keeping an audit trail while holding no model or business logic; the model provider and agent loop are yours to choose. If you pass credentials or sensitive internal services through agents, this is a genuine alternative to trusting the model to behave.