Atlas — Features & Architecture

Architecture

Seven layers, one assistant

Every command runs through a routing pipeline designed to answer fast when it can, and escalate to something smarter when it needs to.

Fast keyword layer

Instant response for known commands — opening apps, wake words, pause/cancel. No model call needed.

~instant

phi3:mini

Handles simple conversational questions directly. Escalates anything that looks like a real task.

local · 2–4s

Mistral 7B

Orchestrates complex multi-step tasks — generates a structured JSON execution plan, speaks a summary, and waits for confirmation before doing anything.

local

DeepSeek Coder 6.7B

Dedicated code generation model — handles every code file Atlas writes, in Python, JavaScript, HTML, and more.

local

Cloud APIs — opt-in

Claude and Gemini are available for long-context reasoning and real-time information — always disabled by default, always asking permission before use.

opt-in only

LLaVA

Local vision model — analyzes webcam frames for object identification, scene description, and reading text in view.

local

MemPalace

Controlled memory recall and formatting — semantic search over what you've told Atlas to remember, injected as context when relevant.

local

The interface

An orb that shows its state

The GUI is a PyQt5 window with a vispy 3D particle orb — per-particle color variation, depth-based sizing, and beam connections between nearby particles for a holographic look. Each state below has distinct particle density, breathing speed, and connection density.

State	Color	Animation
Listening	Blue	Slow, gentle pulse
Thinking	Green	Medium pulse, more particles
Speaking	Light blue	Fast pulse
Sleeping	Yellow	Very slow, deep breath, sparse particles
Error	Red	Rapid jitter

Talk to it

A sample of what you can say

A taste of the full command set — memory, apps, code generation, web, and calendar/email are all voice- or text-driven.

Everyday questions

"how long is the Nile River"	Answered instantly by phi3 — no escalation needed
"what's the capital of South Dakota"	Quick factual answer, 2–4 seconds
"what's 1 plus 1"	Handled directly — simple math and trivia never touch the orchestrator

Memory

"remember that I prefer async python over threading"	Stores preference in long-term memory
"what do you remember about me"	Retrieves stored preferences and facts
"forget that I like X"	Removes a specific stored memory

Code generation

"create a python file called calculator.py with add and subtract methods"	Generates a working Python class
"create a homepage.html with about and contact sections using inline css"	Generates a styled HTML page

Web & browser

"google latest AI news"	Opens browser, searches Google
"navigate to github.com"	Opens URL directly
"click first result"	Clicks the first search result

Calendar & Gmail

"what do I have today"	Reads today's events
"add meeting friday at 2pm for 2 hours"	Creates the event, with duration parsed directly
"reply to {name}"	Drafts a generated response for review

Vision

"what do you see"	Describes the full scene via webcam
"read this"	Transcribes visible text
"is anyone there"	Detects people in frame

First principles

Local by default, cloud by permission

What that actually means

All processing is local by default — nothing leaves your machine unless you say so
Cloud API calls (Claude, Gemini) are disabled out of the box
When enabled, Atlas asks permission before every single API call — not just once at setup
Anthropic and Google don't use API calls to train their models by default

Compatibility

Platform support

Feature	Linux	Mac	Windows
Core AI / LLM	✅	✅	✅
Voice I/O	✅	✅	⚠️
File / code tools	✅	✅	✅
App launching	✅	⚠️	⚠️
Browser control	✅	✅	✅
GUI	✅	✅	✅

⚠️ = works with minor path configuration.
Mac: app paths are Linux-style by default — swap to /Applications/... and replace xdg-open with open.
Windows: not recommended natively — a VM is the smoother path for now.