Aulinx

AI-native Linux. Desktop to server.

Other AI agents look at your screen. Aulinx IS the screen.

What is it? | How it works | Getting started | Compositor | Tools | Roadmap

---

What is Aulinx?

Aulinx is an AI layer for Linux that works at three levels:

┌─────────────────────────────────────────────────────────┐
│  Tier 3: Aulinx Compositor                              │
│  Custom Wayland compositor with semantic scene graph     │
│  AI sees every pixel because it rendered them            │
├─────────────────────────────────────────────────────────┤
│  Tier 2: Aulinx Desktop                                 │
│  AT-SPI GUI control on any desktop (GNOME/KDE/Sway)     │
│  Click buttons, read menus, type text — semantically     │
├─────────────────────────────────────────────────────────┤
│  Tier 1: Aulinx Core                                    │
│  Files, git, process, network, docker, system, packages  │
│  Works headless — servers, WSL, Docker, SSH              │
└─────────────────────────────────────────────────────────┘

187 tools across all tiers. A Wayland compositor (Rust) with 33 IPC commands. Semantic desktop understanding — not screenshots, not OCR.

aulinx > why is my computer slow right now?

  > process_list(sort_by=cpu)

  ┌─ Result (9ms) ────────────────────────────────────┐
  │ firefox (42% CPU), code (18% CPU), slack (8% CPU)  │
  └────────────────────────────────────────────────────┘

  Firefox is consuming 42% of your CPU with 47 tabs open.
  Want me to kill background processes?

aulinx > search for "wayland compositor" in Firefox

  > atspi_set_text(app_name=firefox, element_name=Search, text=wayland compositor)

  ┌─ Result (40ms) ──────────────────────────────────┐
  │ "Set text on 'Search': 'wayland compositor'"      │
  └────────────────────────────────────────────────────┘

Unlike other AI desktop agents that use screenshots, Aulinx reads the actual UI structure — semantic, not pixel-based. No OCR needed.

How It Works

Aulinx has two deployment modes:

Mode 1: Agent on any desktop          Mode 2: Full AI compositor
(works today on GNOME/KDE/Sway)       (custom Wayland compositor)

┌─────────────────────────────┐      ┌─────────────────────────────┐
│  CLI / Web UI / Voice / MCP │      │  CLI / Web UI / Voice / MCP │
├─────────────────────────────┤      ├─────────────────────────────┤
│  Agent (187 tools + LLM)    │      │  Agent (187 tools + LLM)    │
├─────────────────────────────┤      ├─────────────────────────────┤
│  aulinx-semanticd (daemon)  │      │  aulinx-compositor (Rust)   │
│  AT-SPI → Scene Graph → IPC │      │  Smithay + Scene Graph      │
├─────────────────────────────┤      │  + Input Injection + IPC    │
│  GNOME / KDE / Sway / Xfce  │      │  Wayland compositor IS the  │
│  Your existing desktop       │      │  AI-native desktop          │
└─────────────────────────────┘      └─────────────────────────────┘

The scene graph is the key abstraction — a structured representation of every window, UI element, and action on your desktop. Both modes expose the same IPC protocol:

{"method": "scene.windows"}         // list all windows with metadata
{"method": "scene.find", "params": {"role": "button", "name": "Save"}}
{"method": "input.type", "params": {"text": "hello"}}
{"method": "input.key", "params": {"combo": "ctrl+s"}}
{"method": "window.focus", "params": {"window_id": 1}}
{"method": "window.close", "params": {"window_id": 1}}

Getting Started

Mode 1: Agent on your existing desktop

Works on any Linux desktop (GNOME, KDE, Sway, Xfce). No custom compositor needed.

Prerequisites

• Linux with a running desktop (Wayland or X11)
• Python 3.10+
• Ollama with a model that supports tool calling
• python3-pyatspi for GUI control (apt install python3-pyatspi)

Install

git clone https://github.com/aulinx/aulinx.git
cd aulinx
pip install -e .
ollama pull qwen2.5:14b

Run

# Interactive mode
aulinx

# One-shot command
aulinx -c "what windows do I have open?"

# Use a specific model
aulinx -m qwen2.5:14b

# Start the web UI
aulinx --serve
cd ui && npm install && npm run dev
# Open http://localhost:5173

# Resume last conversation
aulinx --resume

# Background daemon with global hotkey (Super+Space)
aulinx --daemon

# Voice input mode (requires faster-whisper)
aulinx --voice

# MCP server for Claude Desktop
aulinx --mcp

# Check system dependencies
aulinx --doctor

Docker (test with a full desktop)

docker compose -f docker/docker-compose.yml up
# Open http://localhost:6080/vnc.html (password: aulinx)
# Inside container: aulinx -m qwen2.5:14b --base-url http://host.docker.internal:11434

Mode 2: AI compositor (Rust)

The custom Wayland compositor with built-in AI understanding.

Prerequisites

• Linux with Wayland support
• Rust 1.75+ (nightly recommended)
• System libs: libwayland-dev libinput-dev libudev-dev libgbm-dev libxkbcommon-dev libseat-dev

Build

cd compositor
cargo build -p aulinx-compositor -p aulinx-semanticd

Run

# Inside an existing Wayland session (opens as a window)
WAYLAND_DISPLAY=wayland-0 ./target/debug/aulinx-compositor

# Launch apps inside the compositor
WAYLAND_DISPLAY=wayland-1 foot    # terminal
WAYLAND_DISPLAY=wayland-1 firefox # browser

# Connect an AI agent via IPC
python3 test_client.py

Compositor

The Aulinx compositor is a Wayland compositor built on Smithay with a semantic scene graph baked in. Every window, UI element, and layout change is exposed as structured data over a Unix socket IPC.

What makes it different

Feature	Traditional WM	Aulinx Compositor
Window info	EWMH/IPC hacks	Semantic scene graph
UI elements	Not accessible	AT-SPI bridge built-in
AI input	xdotool/ydotool	Native keyboard injection
Events	Poll-based	Push subscriptions
Data format	Mixed protocols	Single JSON-RPC API

IPC protocol

# Connect to the compositor's IPC socket
# Default: $XDG_RUNTIME_DIR/aulinx/semantic.sock

# Query windows
echo '{"jsonrpc":"2.0","id":1,"method":"scene.windows","params":{}}' | \
  socat - UNIX-CONNECT:$XDG_RUNTIME_DIR/aulinx/semantic.sock

# Response:
# {"result": [{"id": 1, "app_id": "foot", "title": "foot", 
#   "geometry": {"x": 0, "y": 0, "width": 1280, "height": 800}}]}

# Inject text into the focused window
echo '{"jsonrpc":"2.0","id":2,"method":"input.type","params":{"text":"hello"}}' | ...

# Inject key combos
echo '{"jsonrpc":"2.0","id":3,"method":"input.key","params":{"combo":"ctrl+s"}}' | ...

# Focus a window
echo '{"jsonrpc":"2.0","id":4,"method":"window.focus","params":{"window_id":1}}' | ...

# Close a window
echo '{"jsonrpc":"2.0","id":5,"method":"window.close","params":{"window_id":1}}' | ...

# Subscribe to window events (open/close/focus)
echo '{"jsonrpc":"2.0","id":6,"method":"scene.subscribe","params":{"filter":"*"}}' | ...

Architecture

aulinx-compositor (Rust, ~7,900 LOC)
├── Smithay Wayland compositor (winit + DRM backends)
├── Semantic bridge (window → scene graph sync)
├── IPC server (JSON-RPC over Unix socket)
├── Input injection (xkbcommon keymap → keyboard events)
└── Tiling layout (equal-width horizontal split)

aulinx-semantic (Rust library)
├── Scene graph (windows, elements, actions)
├── AT-SPI source (reads GNOME/KDE UI trees)
├── Direct source (compositor integration)
├── Diff engine (push events on changes)
└── Query engine (scene.windows, scene.find, etc.)

Tools

187 tools across 43 modules. Selected highlights below — run aulinx --doctor or /tools for the full list:

Category	Tools	Count
Window	list, get_focused	2
AT-SPI	get_tree, find_elements, do_action, read_text, set_text, screenshot	6
Files	read, write, edit, move, trash, list, search	7
Text	count, grep, replace, head, tail	5
Git	status, log, diff, commit, branch, stash	6
Apps	launch, list_running	2
Process	list, kill	2
Services	list, status, start, stop, restart	5
Network	status, wifi_list, wifi_connect, wifi_disconnect	4
Audio	get_volume, set_volume, mute	3
Display	list, brightness	2
Power	status, profile, suspend, shutdown	4
Theme	get, set_dark, wallpaper_set	3
Bluetooth	status, scan, connect, disconnect, toggle	5
Input	key_combo, type_text	2
Session	who_am_i, uptime, disk_usage, env_get	4
Packages	search, install, list_installed	3
XDG	open, default_app_get, default_app_set, mime_type_of	4
Timer	set_timer, cancel_timer, list_timers	3
Clipboard	get, set	2
Notifications	send	1
Memory	store, get, delete, list_namespaces	4
D-Bus	list_services, introspect, call	3
OCR	screenshot_ocr, image_ocr	2
DateTime	now, convert, calendar_show	3
System	info, shell_exec	2
Workflow	context_get, wait, audit_recent	3
Workflows	create, list, run, delete, toggle	5
Long Memory	remember, recall, recall_recent, forget, memory_count	5
Server	journal_logs, docker_ps, docker_logs, port_list, firewall_status, cron_list, disk_health, system_logs_summary	8
Compositor	summary, describe, ascii, suggest, status, config, ping, windows, focused, find_window, element_at, screenshot, annotated_screenshot, window_count, type, key, click, drag, scroll, spawn, focus, close, minimize, swap_master, set_ratio, set_gap, batch, diff, wait_for, run_and_type	30

The table lists representative tools per category; remaining tools (clipboard, archive, calc, schedule, sysadmin, productivity, AI, and more) bring the total to 187.

Permission Tiers

Tier	Behavior
Read	Always auto-allowed
Low-risk	Auto-allowed, logged
Mutate	Confirms first time per session, then auto
Destructive	Always confirms
Irreversible	Always confirms with extra warning

Slash Commands

/tools    - List all available tools
/context  - Show current desktop context
/history  - Browse past conversation sessions
/audit    - Show recent tool calls with timing
/doctor   - Check system dependencies
/clear    - Clear conversation history
/help     - Show help

Configuration

Config at ~/.config/aulinx/config.toml (auto-created on first run):

[llm]
model = "qwen2.5:14b"
base_url = "http://localhost:11434"
temperature = 0.3

[permissions]
# Override tool permission tiers
# shell_exec = "mutate"  # uncomment to lower confirmation requirement

Roadmap

Released:

• v0.1–v0.3: 92→103 tools + CLI + web UI + tests + audit + long-term memory + daemon + voice + MCP + plugins
• v0.4.0: Semantic compositor — Wayland compositor with scene graph, 20 IPC commands, input injection, DRM/udev backend
• v0.5.0 (current): Multi-provider LLM (Ollama/OpenAI/Anthropic/Gemini/Qwen), ReAct planner, error recovery, OSWorld benchmark harness, hybrid perception, action grounding, dynamic tool selection, task decomposition, sandboxed execution, history summarization, learning from outcomes, multi-agent delegation, Python SDK, autonomous mode, portal-first screen capture

Planned:

• v1.0: Daily-drivable compositor, full OSWorld-Verified benchmark run, cross-platform stubs, one-command install

See CHANGELOG.md for the detailed per-release history.

Name

Au (gold, element 79) + linx (Linux / lynx). The gold standard of AI-powered Linux.

Contributing

See CONTRIBUTING.md for development setup, code style, and how to add new tools.

# Python agent
pip install -e ".[dev]"
make test   # run tests
make lint   # check code style

# Rust compositor
cd compositor
cargo build
cargo test

License

MIT