Open Source AI Client

Desktop and web client for OpenAI-compatible and Anthropic endpoints. Tauri shell, Vite + React frontend, data stored locally.

Highlights

• Any OpenAI or Anthropic compatible endpoint: no vendor lock-in.
• Runs local models: point it at Ollama or LM Studio and chat without a cloud account.
• Reads pages you link: the model can fetch a URL and answer from the page itself.
• Small and fast: Rust over the OS webview with no bundled browser engine, fonts or icons. ~21MB on disk.
• Secure and private: conversations and API keys stay on your machine, and the app has no analytics.
• Remembers you: durable notes the model writes as it works, plus your own, sent with every chat.
• Updates itself: the desktop build finds a newer release, downloads it and relaunches, and a Settings section checks on demand.
• One codebase, two targets: the same app ships as a desktop binary and as the website.
• Extensible with plugins: add UI, commands, model tools, wire protocols, themes and more from a plugin. Plugins are full-trust — they run in this app's own process with its privileges — so install only ones you trust; the plugin API is unstable at 0.1 and may change between releases (Writing a plugin).
• Completely open source: GPL-3.0-only, you are free to make copies as long as they are open source.

Download

Releases publishes macOS (universal), Windows and Linux builds. The web build runs at aiclient.jyz.land.

macOS installs with Homebrew, and updates itself from then on:

brew trust jeffreyjyz/tap
brew install --cask jeffreyjyz/tap/oss-ai-client

Installing the .dmg by hand instead? Builds are unsigned, so clear the download quarantine flag after installing, or the app will not open:

xattr -dr com.apple.quarantine "/Applications/oss-ai-client.app"

On Windows, SmartScreen warns: pick More info → Run anyway.

Features

• Providers: OpenAI (Responses or Chat Completions), OpenRouter, OpenCode Zen, OpenCode Go, Command Code, Volcengine ARK (platform or Agent Plan), Anthropic, Ollama, LM Studio, or any OpenAI-compatible base URL. OpenRouter's app-identification headers can be switched off per provider.
• Streaming, and a collapsible block for reasoning models' thinking.
• Web search: the endpoint's own tool where it has one (OpenAI, OpenRouter, Anthropic), or TinyFish with your own key where it has none. That is how a local model gets search at all.
• Per-conversation system prompts, and profiles bundling provider, model and prompt.
• File and image attachments.
• Export/Import moves settings and conversations between the desktop and web builds.
• Plugins (off by default): contribute UI slots, commands, model tools, wire protocols, request/stream pipeline stages, themes, settings sections and renderers. A plugin runs in-process and unsandboxed by design — plugin-data is the only capability enforced for real (path-scoped), and net cannot be.

Compatibility

Stored settings carry a dataVersion and a migration registry. Traversal is forward-only: a newer build reads and migrates data written by an older one — a blob with no dataVersion is lifted to the current version and no field is dropped — and conversations live in their own store, never rewritten by a settings read.

Downgrading is tolerated, not migrated. A 0.2.3 install opening data written by 0.3.0 keeps working: the added keys (dataVersion, pluginsEnabled, plugins) are additive and ignored by the older schema, and its conversations survive. It will not know about plugin state, and a provider pointed at a plugin-registered protocol is not usable there. Prefer moving forward.

Development

Needs Bun and a Rust toolchain.

cd web
bun install
bun run dev        # web
bun run tauri dev  # desktop

Gate: cd web && bun run verify (tsc, Biome, the Effect language service, bun test).

License

GPL-3.0-only. The 0.1.x releases were published under MIT and remain MIT.