Run AI Agents Inside Hardware-Isolated VMs
A3S is a modular Rust ecosystem for production AI agents — hardware TEE isolation, privacy-aware security, and a full agent framework in one coherent stack.
The problems A3S solves
Most agent frameworks ignore the hard parts. A3S is built around them.
Hardware isolation by default
Every agent runs inside a MicroVM backed by AMD SEV-SNP or Intel TDX. The host cannot read agent memory. Attestation is built in, not bolted on.
Privacy-aware at every layer
SafeClaw classifies PII, tracks data taint through tool calls, detects prompt injection, and audits every interaction — before it reaches the LLM.
Production-grade performance
Lane delivers 33k–50k ops/sec priority scheduling. Box cold-starts VMs in ~200ms. Power runs local LLMs with RA-TLS attestation. All in pure Rust.
Composable, not monolithic
Use Code standalone, add Box for isolation, plug in SafeClaw for security. Every module has a trait interface — swap any piece without touching the rest.
Build agents, not boilerplate
Infrastructure that runs agents safely
VM-isolated sandbox — OCI images, Docker CLI, WarmPool, CRI, TEE support.
Privacy-preserving LLM inference — OpenAI-compatible API, TEE support, multi-backend, multi-GPU, tool calling.
15 middlewares, Knative autoscaling, HCL hot-reload, TLS/ACME.
PII classification, taint tracking, injection detection, compliance audit (HIPAA / PCI-DSS / GDPR).
Composable primitives
6 built-in lanes, retry/DLQ, rate limiting, persistent storage, 33k–50k ops/sec.
Pluggable event bus — NATS JetStream + in-memory, schema validation, AES-256-GCM encryption.
9-engine consensus ranking — DuckDuckGo, Brave, Google, Baidu, proxy rotation.
Episodic / semantic / procedural / working memory, atomic persistence, relevance decay scoring.
Quick Install
Available on crates.io, PyPI, npm, and Homebrew
cargo add a3s-code-corepip install a3s-code a3s-lane a3s-searchnpm install @a3s-lab/code @a3s-lab/lanebrew tap a3s-lab/tap && brew install a3s-search a3s-powerReady to build?
Start with the Code framework and add modules as you need them.