Axon OS is an AI-native desktop operating system built on Ubuntu 24.04 LTS with GNOME Shell. The system is designed around two centralized D-Bus services — Axon Brain and Axon Context — that provide a shared AI backbone for all desktop components. All inference runs locally through Ollama, keeping user data entirely on-device.
┌─────────────────────────────────────────────────────────────┐
│ User │
└─────────────────────────┬───────────────────────────────────┘
│
┌─────────────▼────────────────────────┐
│ GNOME Shell │
│ │
│ ┌──────────┐ ┌──────────────────┐ │
│ │Intent Bar│ │ AI Panel │ │
│ └────┬─────┘ └───┬──────────────┘ │
│ │ │ │
│ ┌────▼────────────▼────┐ ┌──────┐ │
│ │ Space Manager │ │ Dock │ │
│ │ (AI auto-routing) │ │ │ │
│ └──────────┬──────────┘ └──┬───┘ │
└─────────────┼────────────────┼──────┘
│ D-Bus │
┌────────────────┼────────────────┼──────────────┐
│ │ │ │
┌────────▼─────────┐ ┌───▼──────────┐ ┌───▼───────────┐ │
│ org.axonos.Brain │ │ org.axonos. │ │ AI-Native │ │
│ │ │ Context │ │ Applications │ │
│ ├ Generate() │ │ │ │ │ │
│ ├ SendMessage() │ │ ├ Window │ │ ├ Terminal │ │
│ ├ ClassifyIntent │ │ │ tracking │ │ ├ Files │ │
│ ├ ClassifyWindow │ │ ├ Clipboard │ │ └ Settings │ │
│ ├ PullModel() │ │ │ hooks │ └───────────────┘ │
│ ├ ListModels() │ │ ├ Open file │ │
│ ├ GetStatus() │ │ │ scanning │ │
│ └ Conversations │ │ └ Terminal │ │
│ (SQLite CRUD) │ │ history │ │
└────────┬─────────┘ └─────────────┘ │
│ HTTP (localhost:11434) │
┌────────▼─────────┐ │
│ Ollama Daemon │◄───────────────────────────────────┘
│ (systemd user) │
└────────┬─────────┘
│ llama.cpp / CUDA / ROCm
┌────────▼─────────┐
│ GPU / CPU │
└──────────────────┘
The Brain is the central AI gateway for the entire operating system. It is a D-Bus session service that:
- Wraps Ollama: All components communicate with the Brain via D-Bus; only the Brain talks to Ollama over HTTP. This provides a single point of model management, caching, and error handling.
- Routes tasks to models: The Brain maps different task types to different model tiers:
- Speed model (e.g.,
llama3.2:1b): Used for instant classifications — window routing, intent parsing. - General model (e.g.,
llama3.2:3b): Used for chat, summarisation, everyday queries. - Deep model (e.g.,
llama3:8b): Used for complex reasoning, code generation, deep analysis.
- Speed model (e.g.,
- Profiles hardware: On first launch, the hardware_profiler.py scans system RAM, GPU vendor (NVIDIA, AMD, Intel), and VRAM to recommend appropriate models for each tier.
- Persists conversations: Uses SQLite (
~/.axon/conversations.db) to store chat history with full CRUD operations. - Streams tokens: Supports both synchronous and streaming generation via D-Bus signals (
TokenGenerated,GenerationCompleted).
| D-Bus Method | Signature | Purpose |
|---|---|---|
GetStatus() |
→ s |
JSON status: Ollama online, active models, config |
ListModels() |
→ s |
JSON array of locally pulled models |
PullModel(name) |
s → b |
Starts background model download, emits PullProgress signals |
Generate(prompt, context, model, stream) |
sssb → s |
Unified text generation |
CreateConversation(system_prompt, title) |
ss → s |
Creates a new conversation, returns ID |
SendMessage(conv_id, message, context, model, stream) |
ssssb → s |
Sends a message in a conversation |
ClassifyIntent(text) |
s → s |
Classifies natural language into system actions |
ClassifyWindow(title, wm_class) |
ss → s |
Routes a window to a named workspace |
ClassifyWindowDetailed(title, wm_class) |
ss → s |
JSON {space, confidence, source}; the Shell asks the user when confidence is low |
ArchiveConversation(conv_id, archived) |
sb → b |
Hides or restores a conversation in the AI panel list; False if it doesn't exist |
The Context Engine is an ambient intelligence service that aggregates user activity into a structured JSON snapshot:
- Window tracking: Monitors the active window title, WM class, and position via
Wnck. - Workspace tracking: Tracks the current named Space.
- Clipboard hooks: Optionally captures clipboard contents.
- File scanning: Detects open files in code editors by inspecting
/proc/*/fd. - Terminal history: Reads recent shell commands from history files.
| D-Bus Method | Signature | Purpose |
|---|---|---|
GetActiveContext() |
→ s |
Full JSON snapshot of current desktop state |
GetContextString() |
→ s |
Pre-formatted, prompt-ready context string |
Spaces are persistent, named workspaces with AI-powered window auto-routing. When a new window opens, the Shell extension calls org.axonos.Brain.ClassifyWindow(title, wm_class) to determine which Space it belongs to. The nine default spaces are:
| Space | Typical windows |
|---|---|
| Code | VS Code, JetBrains, vim |
| Web | Firefox, Chrome |
| Chat | Discord, Telegram, Slack |
| Files | Nautilus, file managers |
| Media | VLC, Spotify, GIMP |
| Work | LibreOffice, Google Docs |
| Personal | Email, calendar |
| Terminal | Terminal emulators |
| Notes | Obsidian, Notion, text editors |
The GNOME Shell extension owns this name on the session bus while it is enabled, at
/org/axonos/Shell.
| D-Bus Method | Signature | Purpose |
|---|---|---|
SwitchToSpace(index) |
i → |
Switches to Space index (0-based), as Super+1…9 does |
User-facing strings use the axon-os gettext domain. Python code marks them with
_() from services/i18n.py (N_() for constants that code compares); the Shell
extension uses _t() from ui-logic.js, which extension.js connects to the
extension's gettext. Both use {name} placeholders, so shared strings have one msgid.
scripts/update-pot.sh refreshes po/axon-os.pot, and tests/test_po_template.py
fails when a marked string is missing from it.
The Intent Bar (apps/intent-bar/) is a keyboard-driven command palette activated via Super+Space. It accepts natural-language intents and routes them through org.axonos.Brain.ClassifyIntent():
- Run command:
{"action": "run_command", "command": "sudo apt update"} - Open application:
{"action": "open_app", "app": "firefox"} - Default answer: Plain text response displayed inline.
The AI Panel (apps/axon-ai-panel/) is a persistent sidebar that provides a conversational interface. It:
- Calls
org.axonos.Brain.SendMessage()with streaming enabled. - Includes desktop context from
org.axonos.Context.GetContextString()in every query. - Maintains persistent conversation history via the Brain's SQLite store.
| Path | Purpose |
|---|---|
~/.axon/config.json |
Model configuration (speed, general, deep model names) |
~/.axon/conversations.db |
SQLite database for chat history |
~/.axon/hardware_profile.json |
Cached hardware profiling results |
~/.axon/spaces.json |
Persisted Space state |
~/.axon/logs/ |
Application and service debug logs |
- User types in Intent Bar → Intent Bar calls
Brain.ClassifyIntent(text)over D-Bus. - Brain receives request → Routes to Speed model for fast classification → Returns JSON action or plain text.
- User opens AI Panel → Panel calls
Context.GetContextString()to gather desktop state → CallsBrain.SendMessage()with context + user message → Brain streams tokens back viaTokenGeneratedsignal. - User opens a new window → Shell extension detects
window-createdsignal → CallsBrain.ClassifyWindow(title, wm_class)→ Moves window to the appropriate Space.