ATLAS OS

Inspiration

Most AI tools today are reactive — you ask, they answer. We wanted to build something fundamentally different: an AI that thinks ahead, acts autonomously, and works across your entire digital life without you having to babysit it. We were inspired by the idea of a true AI operating system — one that combines a proactive personal agent for individuals with an enterprise-grade intelligence layer for companies.

The name ATLAS reflects our ambition: a system that can carry the cognitive load across research, code, communication, and business operations simultaneously.

What it does

ATLAS OS is two products in one.

The first is a personal AI agent that runs as a persistent desktop app on macOS and Windows. It can autonomously use your computer — clicking, typing, navigating apps on your behalf. It does deep research across up to 500 sources, edits video, handles voice interactions, and connects to Telegram, Discord, Slack, WhatsApp, Matrix, and email. It has 70+ memory subsystems including episodic, semantic, emotional, and knowledge graph memory, and a Skill Forge that lets it synthesize and sandbox new tools on demand.

The second is an enterprise brain service. It ingests signals from Slack, GitHub, and Jira, matches them against natural language rules, runs an LLM judge to evaluate flagged events, and proposes governed safe-write actions that go through a human approval flow with a full audit trail. Everything is isolated per tenant using Postgres row-level security.

How we built it

The stack splits across three layers. The agent core is Node.js and TypeScript (around 784 files) with a native Rust performance layer (around 88 files) for SIMD vector operations that give us a 15–20x speedup on embedding search. All LLM calls proxy through Convex with end-to-end encryption via ECDH P-256 and AES-256-GCM. A native model router picks between Claude, GPT-5, and Gemini based on task complexity and iteration depth. Memory runs on local SQLite with HNSW ANN indexing, a knowledge graph, 6-hour decay cycles, and nightly dream replay consolidation. The enterprise layer is a standalone Node service on Railway with Postgres and pgvector on Neon, 34 forward-only migrations, and a cryptographic hash chain for audit compliance. Dynamically synthesized tools run inside OS-level sandboxes — macOS sandbox-exec, Linux seccomp with Landlock and network namespace isolation — supervised by a Rust napi module.

Challenges we ran into

Getting the AI to reliably click UI elements at scale required building a 4-tier cascade: accessibility tree, then OmniParser, then a local vision model (UI-TARS), then cloud grounding — each with fallbacks and self-correction. Building truly fail-closed multi-tenant security required extensive testing; our auth tests are literally security tests. We also discovered that many obvious Rust ports were actually slower due to FFI crossing costs, so we documented the lessons carefully.

Accomplishments we're proud of

We shipped a complete, production-deployed enterprise SaaS with end-to-end encryption, tenant isolation, governed action execution, and a cryptographic audit ledger — all in one codebase. We built a Skill Forge that lets ATLAS synthesize its own tools: the LLM writes code and tests, a sandbox runs them, SAST critics check them, and the system retries up to 10 times. We have a four-tier agent orchestration system going from single sub-agents up to long-running tasks with 200-iteration phases and checkpoint resume.

What we learned

The performance boundary between TypeScript and Rust is not where most people think — Node.js hardware-accelerated crypto is 2–4x faster than software Rust. Time your TS code first; port to Rust only when you have actual proof. Multi-agent systems need explicit handoff protocols to maintain quality across agent boundaries. And enterprise AI security is fundamentally about audit trails, not just access control.