SYSTEMS · AI-NATIVE OS

Rimuru OS

A local-only, AI-native Linux distribution where you control the desktop by describing what you want instead of clicking through it.

↓ SCROLL

The problem

Every desktop OS still assumes you drive it by hand: hunting through menus, memorizing terminal flags, and repeating the same setup steps over and over. AI sits bolted on as a chat box in the corner, unable to actually touch the system. Rimuru starts from the opposite premise, that the AI is the interface. You boot the machine, say 'set up my development environment for Rust,' and the system parses the intent, plans the steps, and executes them across the terminal, GUI apps, and browser. The hard part is doing this reliably and privately, without shipping your screen to a cloud API.

The approach

Rimuru runs entirely on the local machine using quantized models served through Ollama, so nothing leaves the device. Natural language from voice or a search panel is routed through an orchestration layer that parses intent, plans a goal, and picks the right model for the job: a fast 3B model for reactions, an 8B for multi-step planning, a 14B for hard problems, and a vision model for reading the screen. To actually control applications it uses a four-tier fallback chain, always preferring the most reliable method available: native APIs first, then the accessibility tree, then pre-recorded OSWorld interaction patterns, and only as a last resort vision plus simulated input. Every autonomous action is gated by confidence, checkpointed for undo, and written to an append-only audit log.

System design

Base systemArch Linux base, Wayland display server, systemd, zsh
Core languagesRust (system/performance), Python (AI/ML), Go (networking)
AI engineOllama and llama.cpp serving quantized local models (3B–14B + LLaVA vision)
Data storeChromaDB for vector storage and long-term memory
PerceptionOpenCV, YOLOv8 UI detection, easyOCR, AT-SPI2 accessibility (pyatspi2)
ControlNative app APIs, AT-SPI2 tree, OSWorld patterns, Playwright/Selenium, pexpect
UI layerGTK4 overlay panels, Rofi/Ulauncher launcher, custom voice indicator widgets

How it works

01

Capture intent

You speak or type a request through a voice indicator or search panel, for example 'set up my Rust environment.'

02

Orchestrate

The orchestration layer parses the intent, builds a goal plan, and routes each step to the right local model: reactive 3B, reasoning 8B, deep-planning 14B, or the LLaVA vision model for reading the screen.

03

Ground the UI

Perception combines screen capture, OCR, YOLOv8 detection, and the AT-SPI2 accessibility tree to locate the exact elements to act on, using Set-of-Marks overlays so the vision model can point at numbered targets.

04

Execute via fallback chain

The executor tries the most reliable control method first: native API, then accessibility tree, then a recorded OSWorld pattern, then vision plus input simulation, acting across terminal, GUI, and browser.

05

Gate and checkpoint

Confidence-based execution runs low-risk tasks automatically, uses countdown timers for medium confidence, and asks for approval otherwise. A checkpoint is taken before every change so any action can be rolled back.

06

Log and learn

Every action lands in an append-only audit log, and recorded user trajectories let the system learn to replay tasks by semantic matching rather than pixel coordinates.

What it does

Natural-language control

Drive the terminal, desktop apps, and browser by describing the outcome instead of performing each click.

Fully local, private by default

All inference runs on-device through Ollama and llama.cpp, so no screen data or prompts leave the machine.

Four-tier reliable automation

A fallback chain from native APIs down to vision plus input simulation keeps control working even when an app exposes no clean interface.

Trajectory recording and replay

Demonstrate a task once and the system learns to repeat it, adapting the recorded pattern to environmental changes through semantic matching.

Undo, sandbox, and audit

Checkpoints enable rollback of file, package, and config changes; sandbox mode previews what an action would do; an append-only log records everything the AI touches.

Confidence-gated autonomy

High-confidence, low-risk actions run on their own while riskier ones require a countdown or explicit approval.

Stack

RustPythonGoOllamallama.cppChromaDBLLaVAOpenCVYOLOv8easyOCRAT-SPI2PlaywrightGTK4WaylandArch Linuxwhisper.cpp
2025Year
SoloRole
Pre-alphaStatus
GPL-3.0License