Dez
In development GitHub

In development — no release yet

Dez: The free and open-source IDE for the local inference-first paradigm

An editor built on the Atlas Inference Engine, WebGPU and Rust compiled to WebAssembly. The model runs in your browser, on your GPU. No server round-trip. No API key. Your code and your weights never leave the machine.

Follow development on GitHub What is Dez?
  • Atlas Inference Engine
  • Rust
  • WebGPU
  • 100% WebAssembly

What it is

An editor that does its own thinking

Dez is an IDE whose language model runs where your files already are. It is built on the Atlas Inference Engine — a pure-Rust LLM inference stack — compiled to WebAssembly and dispatched to your GPU through WebGPU. Open the page, load a model, and the editor works with no account and no network.

The reason to do this is not benchmark bragging. It is that “local-first” stops being a setting you can forget to turn on. There is no inference endpoint to leak to, no key to rotate, and no vendor between you and your own source code.

In development. Dez is an early work in progress. There is nothing to download yet, and no release date. This page exists so you can follow the work as it happens.

Principles

Four commitments, not four features

  • Local-first, not local-optional

    Inference runs on the machine in front of you. Your source, your prompts and your weights are never uploaded, because there is nowhere to upload them to. No account, no API key, no per-token bill.

  • WebGPU for compute

    Dez targets the WebGPU standard rather than a single vendor toolchain, so the same build talks to the GPU already in your laptop — whoever made it.

  • 100% WebAssembly

    The engine compiles to WebAssembly and ships as static files. Nothing is installed, nothing runs as a privileged process, and the browser sandbox stays the trust boundary.

  • Free and open source

    Dez follows Atlas: source-available to read, fork and audit under the AGPL-3.0. A tool you run on your own hardware should be a tool you can inspect.

How it works

Atlas, retargeted at the browser

The pipeline is deliberately short. Every stage below is a component that already exists in the open — the work of Dez is joining them into an editor.

  1. Atlas, in Rust

    Atlas is a pure-Rust LLM inference engine: scheduler, model graph and hardware-specific kernels behind tight trait boundaries, with no Python runtime anywhere in the request path.

  2. Compiled to WebAssembly

    That same Rust core is compiled to WebAssembly instead of a native binary. The scheduling and tokenisation logic is the code Atlas already runs on servers — retargeted, not rewritten.

  3. Dispatched over WebGPU

    Atlas is built around swappable hardware backends. In the browser that backend is WebGPU: compute shaders do the matrix work, and WebAssembly drives them.

  4. The editor is the client

    The IDE talks to the engine in-process. Completion and chat are function calls into WebAssembly, not HTTPS requests to somebody else’s datacentre.

Atlas is open source under the AGPL-3.0 and ships hardware- and model-specific kernels behind swappable backend traits — the property that makes a WebGPU backend a port rather than a rewrite. Read the engine at github.com/Avarok-Cybersecurity/atlas.

There is nothing to download yet

Dez is early. Rather than a fake install button, here is the honest one: watch the engine it is built on, and you will see Dez arrive as it is built.

Watch Atlas on GitHub Join the Discord