Skip to content
vivek3071997Public

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

2 Commits

Folders and files

Repository files navigation

AI-os

A local-model chat desktop, packaged as a Debian Linux live-image project.

AI-os replaces the earlier GPU-only hardware research as the active project. Instead of a launcher grid, taskbar, and desktop icons, the user interacts with one chat window. A local language model interprets requests, selects supported actions, and recommends apps using conversation context and app-request counts.

Status: what is and is not delivered

Component Status
Chat UI and local Ollama adapter Implemented; UI and controlled-response integration tested
Constrained actions, confirmations, app-usage persistence Implemented; preview/API tests verified
Debian 13 amd64 live-image configuration Implemented, but not built or boot-tested here
Actual model inference in this environment Not verified: Ollama and model weights are unavailable
Linux desktop app launching / audio / power here Not verified: development host is Windows
Downloadable ISO / installation image Not produced; Linux build environment is required
CPU-free operation / replacing the Linux kernel with predictions Not supported and not claimed

This is not yet a complete production OS. It retains Linux's deterministic kernel, drivers, memory protection, permissions, filesystem, and scheduling. Removing those algorithms would not produce a reliable operating system. The model decides supported user-facing actions; it does not replace every internal OS operation. Conventional CPU hardware remains necessary.

Experience

  • One chat surface for requests, execution results, errors, and confirmations.
  • Suggested app cards appear in the same window after model responses.
  • App request/launch counts persist and are sent to the model for future recommendations. Suggestions are model-selected, not hardcoded rankings.
  • No cloud provider, API key, CDN, analytics, or external UI framework.
  • Local model setup appears inside the chat shell.
  • Power operations require an explicit, expiring, single-use confirmation.
  • Apps open native windows. The chat is the desktop shell, not a replacement rendering engine for browsers, editors, or other applications.
  • On the packaged desktop, Super+Space returns to chat; Alt+Tab switches windows. There is no desktop panel or conventional app launcher.

The shell uses plain HTML/CSS/JavaScript and Node.js built-ins, with no npm dependencies. Chromium, native apps, Linux, and the model runtime still have substantial resource costs. No minimum footprint, speed, or power target has been measured. Model weights will generally dominate image size.

Run the application now

Requires Node.js 20+ and a separately installed local Ollama runtime/model:

npm test
npm start

Open http://127.0.0.1:3210. The default is preview mode: OS-changing actions are explicitly simulated. Real system-status reads still work.

On a machine with Ollama installed, provision a small local model, for example:

ollama pull qwen2.5:3b

Select it in the UI. Model compatibility/plan quality must be evaluated on the chosen model; this example is not an accuracy guarantee. Start the Ollama service with OLLAMA_NO_CLOUD=1. AI-os does not download weights automatically and will not silently substitute keyword routing if inference fails.

For real desktop actions on an existing Linux graphical session:

AIOS_MODE=desktop node src/server.mjs

Install the applications listed in the image package list. The image recipe handles these packages and session environment automatically. Do not run the desktop server as root.

Supported model-selected operations

Action Behavior
Launch an app Fixed installed browser, files, editor, calculator, network settings, or sound settings
System status Memory availability, uptime, platform
Volume Set default audio sink to 0-100% through WirePlumber
Power Suspend, reboot, or power off through systemd after confirmation

Unsupported tasks are reported, not translated into arbitrary shell commands. File deletion, package installation, arbitrary executables, unrestricted scripts, kernel policy prediction, general document automation, and full control of third-party apps are not implemented. An AI response alone is not proof of execution; the UI separately displays action results.

Build the Linux live image

See image build instructions.

The recipe targets Debian 13 (trixie), amd64, not Ubuntu. It packages Linux, Openbox, a Chromium chat shell, a local Ollama service, selected native apps, and your pre-provisioned model. The intended output is a hybrid live ISO.

The build requires a dedicated Linux VM, root privileges, network access to Debian package repositories, and adequate storage for the image and model. This Windows environment has neither a usable Linux build distribution nor a running Docker engine. The attempted build failed before image creation.

There is no disk installer in the recipe. A live image is not the same as an installed OS. Boot and session validation remain necessary before release. No USB drive or physical disk is written by the build script.

Architecture

Chat window
    |
Loopback Node.js service
    |
Local Ollama model -> validated structured action plan + app suggestions
    |
Deterministic capability checks and explicit power confirmations
    |
Unprivileged Linux desktop apps / WirePlumber / systemd
    |
Linux kernel, drivers, CPU, GPU, RAM, storage

The model has no unrestricted shell and does not receive arbitrary executable paths. It chooses from a fixed action schema. Predictions are rejected on invalid fields, unavailable apps, invalid arguments, or malformed output. The backend binds only to 127.0.0.1, restricts Host/Origin, and requires a session token for model/action APIs. Model text is displayed as text, not HTML. These controls do not make a shared local Linux account a security boundary or turn kiosk mode into a hardened appliance.

The model endpoint must be a numeric loopback HTTP origin. Cloud-tagged model names are rejected; the image disables Ollama cloud support. The separately configured Ollama service is still part of the trust boundary. AI-os does not verify that a third-party local server actually performs inference locally.

State and settings

  • Default state: ~/.local/share/ai-os/state.json.
  • Stores selected model and per-app request/launch counts, not chat transcripts.
  • Conversation context remains in memory (last six exchanges) and resets on server restart/model change.
  • History writes use a temporary file and atomic rename; corrupt state errors are surfaced rather than silently discarded.
  • Live-session state disappears on shutdown unless persistence is configured.
Environment variable Default / purpose
AIOS_MODE preview; desktop enables Linux actions
AIOS_PORT 3210
AIOS_DATA_DIR Override the state directory
AIOS_OLLAMA_URL http://127.0.0.1:11434, numeric loopback only
AIOS_MODEL Optional initial model; saved selection takes precedence

Tests and verification

npm test
npm run test:legacy

Tests cover structured model requests, validation, explicit errors, HTTP adapter integration, usage history, confirmation expiration/replay, and same-origin/session checks. Controlled model responses are test fixtures, not real inference quality evidence. Linux-only command execution tests are skipped on Windows.

Browser checks exercise model setup, chatting, app suggestions, literal-text rendering, confirmations, and narrow-screen layout. No production model or Linux ISO boot has been validated here.

Preserved research

Earlier hardware, simulator, and Linux-on-model work is kept under legacy/gpu-platform. It is not part of the new desktop or live-image payload. Existing root fixture/build directories are ignored and are not embedded in AI-os. Legacy instructions must be run from that project's directory; its integration fixtures need to be provisioned there explicitly.

License and repository

Original code is MIT licensed. Debian packages, Ollama, model weights, and legacy Linux fixtures retain their own licenses. Check model redistribution terms before distributing an ISO containing weights.

The configured repository remains vivek3071997/AI-GPU-OS. The project name is now AI-os; the remote repository has not been renamed. These refocus changes have not been pushed.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages