⬡ ARIA — Autonomous Reasoning & Interaction Agent

VISION FOR YOUR COMPUTER

ARIA is a personal AI assistant that natively controls your own devices. Unlike browser extensions or DOM-parsing scripts, ARIA relies on Pure Vision. If a human can see it, ARIA can interact with it. It answers you on the channels you already use (Telegram, Webhook, Email) and renders a visual workspace of its actions.

If you want an autonomous, single-user assistant that feels local, fast, and capable of operating any software on your computer, this is it.

Vision · Architecture · Contributing · NPM Registry

Preferred setup: run aria onboard in your terminal. ARIA guides you step by step through setting up your LLM providers, execution mode, and notification channels.

Models (selection + auth)

ARIA is completely provider-agnostic. We currently support:

• Anthropic (Claude 3.5 Sonnet) - Recommended
• OpenAI (GPT-4o)
• Google (Gemini 1.5 Pro)
• Local/Open-Weights via Ollama (Qwen2-VL, LLaVA)

Install (recommended)

Runtime: Node 22 LTS or newer.

npm install -g @ariacore/cli@latest
# or: pnpm add -g @ariacore/cli@latest

aria onboard

Quick start (TL;DR)

aria onboard

# Start the daemon
aria serve

# Direct CLI usage
aria run "Open YouTube, search for trending music, and take a screenshot"

How it works (short)

Telegram / CLI / Webhook / Email
               │
               ▼
┌───────────────────────────────┐
│          ARIA Gateway         │
│     (orchestration layer)     │
└──────────────┬────────────────┘
               │
               ├─ Vision LLM (Perception)
               ├─ LanceDB (Memory)
               ├─ Safety Gate (HITL)
               └─ Computer Controller (Action)
                         │
                         ▼
             Windows / macOS / Linux

Key subsystems

• Perception-Reasoning-Action (PRA) Loop — The core cycle. ARIA takes a screenshot, reasons about the goal, and executes native input.
• Meta-Cognition Engine — When ARIA gets stuck, it opens a browser, searches the internet, and saves the learned procedure to LanceDB.
• Dual-Memory Store — SQLite for step-by-step episodic logs. LanceDB for vector-based procedural workflows.
• Safety Gate — Human-in-the-loop protection. Destructive bash commands or file deletions require manual CLI/Telegram approval.
• Delivery Transports — Asynchronous reporting via Telegram, Email (Nodemailer), or standard Webhooks.

Development Channels

• stable: tagged releases (v1.x.x), npm dist-tag latest.
• beta: moving head of main, npm dist-tag next.

Security defaults

ARIA connects directly to your operating system. Treat untrusted tasks with extreme caution. By default, any action flagged as destructive will suspend the agent until a user inputs y in the terminal or replies approve on Telegram.

Run aria doctor to surface risky/misconfigured policies in your environment.

Everything we built so far

Core platform

• @ariacore/agent: The PRA loop with checkpointing, error recovery, and tool dispatch.
• @ariacore/computer: The native bridge. node-screenshots for visual capture. xdotool (Linux), osascript (macOS), and PowerShell (Windows) for inputs.
• @ariacore/memory: SQLite event-sourcing and LanceDB embeddings.
• @ariacore/delivery: Markdown generation, HTML artifact export, Telegram polling.

Community

See CONTRIBUTING.md for guidelines, architecture, and how to submit PRs. AI/vibe-coded PRs welcome! 🤖