Desktop app

glia ships a small graphical shell — a chat window that also shows the live glass box (streaming tokens, tool calls, approvals) as it happens. It runs a real glia.Agent under the hood, so it doubles as a demo of the library.

Two ways to get it

1. Download a binary (no Python needed)

Grab the standalone build for your OS from the latest release:

• glia-shell-windows.exe
• glia-shell-macos
• glia-shell-linux

Double-click it (or run it from a terminal). It starts a local server and opens the app in your browser. Nothing is uploaded anywhere — everything runs on your machine.

2. Install from PyPI (native window)

pip install "glia-agents[shell]"
glia-shell

Installed this way, the app opens in a native desktop window (via pywebview), falling back to your browser if a window backend isn't available. Use glia-shell --web to force browser mode.

Using it

• Conversations are saved. The sidebar keeps your chat history (stored as glia trajectory checkpoints in your config directory), so it survives restarts — start new chats, switch between them, or delete them.
• Replies render as Markdown — bold, lists, links, and code blocks with copy buttons. Pick a light/dark theme from the sidebar.
• Works offline out of the box. In demo mode it echoes your messages back, so you can try the interface and watch the event stream with zero setup.
• Run a local model — free. In Settings, choose Ollama, set the model (e.g. qwen2.5 or deepseek-r1) and host, and save. With Ollama installed and the model pulled (ollama pull qwen2.5), everything runs locally on your machine — no key, no network.
• Add a key for real Claude. Open Settings, choose Claude, paste your Anthropic API key (it's stored locally in your config directory and never sent back to the UI), pick a model, and save. Now messages go to Claude, and the two demo tools (current_time, word_count) show up in the glass box when used.
• The glass box on the right lists every event exactly as the library emits it — model_call, model_delta, tool_called, tool_returned, approval_resolved, run_finished. Toggle it with the Glass box button.
• New chat starts a fresh conversation; Quit stops the local server.

How it's built

The shell is deliberately thin and dependency-light:

• The server is pure Python standard library (http.server) — no web framework. Chat replies stream over Server-Sent Events.
• The UI is a single self-contained HTML file (no build step).
• The only optional dependency is pywebview, and only for the native window; the downloadable binaries run in browser mode with zero third-party deps.

In other words, the app is itself a glass box — you can read all of it under glia/shell/.