# The Atlas AI stack

Atlas runs AI on three principles: **route by cost** (free/local first, paid last), **review before write** (no autonomous money/comms/self-modify), and **keep a permanent spine** (every decision and event is logged and linkable).

```mermaid
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#0c1428','primaryTextColor':'#cfe2ff','primaryBorderColor':'#1F6FE0','lineColor':'#4EA0FF','fontFamily':'monospace'}}}%%
flowchart TB
    Senses[Senses · email · news · WhatsApp · bank]:::a --> Spine[(kernel.sqlite · the spine)]:::b
    Spine --> Brain[Atlas Brain · autonomous core]:::a
    Brain --> Gate{Policy gate}:::g
    Gate -->|money / comms / self-modify| Human[Human approval · Control screen]:::h
    Gate -->|safe| Router[LLM router :8888]:::a
    Router --> Local[Local · LLaMA · hermes3 · Qwen]:::c
    Router --> Cloud[Free cloud · Groq · Cerebras · OpenRouter]:::c
    Router --> Claude[Claude · paid · last resort]:::c
    classDef a fill:#0c1428,stroke:#4EA0FF,color:#cfe2ff
    classDef b fill:#145ED2,stroke:#8EC2FF,color:#fff
    classDef c fill:#080c18,stroke:#1F6FE0,color:#9fb8e0
    classDef g fill:#0c1428,stroke:#D4A24C,color:#D4A24C
    classDef h fill:#080c18,stroke:#4EA0FF,color:#cfe2ff
```

## The spine — `kernel.sqlite`

The single source of operational memory. Senses feed it continuously; the Brain reads from it; every decision and event is appended and stays linkable across time. It is the answer to *"why did this happen?"* — one place, queryable, durable.

## The Brain — reviewed execution

The autonomous core reasons over the spine and proposes actions. It **never** writes to money, external communications, or its own configuration without explicit human approval through the Control screen. Daily brief at 07:45; accountability loop morning / midday / evening. A kill-switch lives in the app.

## The router — cost-tiered model selection

A single OpenAI-compatible endpoint (`localhost:8888`) picks the cheapest model that can do the job:

| Tier | Models | Used for |
|---|---|---|
| **Local (free)** | LLaMA 1B · hermes3 · Qwen Coder · Qwen 7B | classification, Dutch copy, code, summaries |
| **Free cloud** | Groq `llama-3.3-70b` · Cerebras · OpenRouter | bulk 70B reasoning |
| **Paid** | Claude (Sonnet / Opus) | multi-step reasoning, architecture — last resort |

The rule: never spend a paid token on a task a free model can handle.

## Context compression

A Headroom proxy sits in front of the Anthropic endpoint to compress context (fewer tokens per call) before requests leave the box — applied transparently to the Brain via `ANTHROPIC_BASE_URL`.

## Where AI surfaces in the OS

- **Atlas Chat** (Open WebUI) — ChatGPT-style UI over the local models or Claude.
- **AI Gateway** — the router's HTTP face.
- **Remote Desktop** — Claude Code runs inside the in-browser Linux desktop; the OAuth login works because the desktop has a real browser.
- **Today / Inbox / Pipeline** — the Brain's briefs, drafts and lead reasoning rendered as cockpit panels.

> Honest status: the spine and router are live; full cross-device chat-history consolidation into the spine is a tracked follow-up, not yet complete.