A new open source chat client called ivx/ai takes a minimalist approach to AI interaction: no accounts, no backend, no telemetry, and no data leaving your browser. The application connects directly to whatever model endpoint you configure, whether that is a local inference server or a cloud API, and keeps all conversations and credentials stored locally.
What the app actually does
The core proposition is straightforward. You point the client at a model running on your own machine or provide an API key for a cloud provider. The application communicates directly with that endpoint. Nothing phones home, not even error reports. API keys can sit encrypted under a user-supplied passphrase, so even if someone accesses the browser storage, the keys remain protected.
The client supports multiple providers and can fall back to WebLLM inference through MLC for browser-based local inference when no external endpoint is available. Chats can be shared anonymously, which is useful for debugging or showing someone a conversation without exposing the full context of your setup.
The interface is designed to be clean and lightweight. In the browser, you can install it as a progressive web app that works offline after the first load. For providers that refuse calls made directly from a browser, a bridge component handles the connection. On macOS, a single brew command installs it. For other platforms, installers are available from the project's releases.
The desktop app and its constraints
The native desktop application uses Tauri, which means it is built with Rust and Node and produces a lightweight binary with a web-based UI. Building from source requires the Tauri prerequisites and a submodule clone of the web interface. The desktop version bundles the same bridge that the web version uses externally, so no additional configuration is needed.
The installers for Windows and Linux are not signed with paid certificates, which means Gatekeeper on macOS and SmartScreen on Windows will flag them on first launch. The release notes document the workaround. This is a common limitation for open source projects that do not have the budget for code signing, but it adds friction for less technical users.
What it does not do
The project has no ads, no paid tier, and no analytics. It does not train on your data because there is nothing to train on. The "bring your own models" approach means the application is just a client, not a platform with its own backend infrastructure.
The planned feature list includes MCP support, tool integration, a store for connectors and skills, chat organization, and browser extensions for Chrome and Firefox. These are not shipping yet, so the current version is a focused chat interface without the bells and whistles that some users might expect.
What this means for developers who care about privacy
The privacy-first design is the main differentiator. Most AI chat clients either route through a proprietary backend or require an account. This one does neither. If you are running a local model through Ollama or a similar tool, the client connects to it without any intermediary. If you are using a cloud API, the key stays in your browser and the request goes straight to the provider.
The GPL v3.0 license means the code is free to use, study, modify, and redistribute, with the requirement that distributed modifications carry the same license. Contributions are welcome, and the project has a Discord for discussion.
For developers who want a clean interface to their models without the overhead of accounts, subscriptions, or telemetry, ivx/ai is a lightweight option that respects the boundary between your browser and the outside world. The desktop app adds convenience for users who prefer a native window, though the unsigned binaries and Tauri build requirements mean it is better suited to developers than casual users at this stage.