Install¶

Pick the path that matches you. A complete beginner on Windows wants the installer; a developer wants pip.

Windows — one-click installer (for everyone)¶

The friendliest path: no Python, no terminal, no configuration.

Download Collie-Setup.exe from the latest release. Windows installers are code-signed through the SignPath Foundation's free code-signing program for open-source projects.
Double-click it. Pick your language on the first screen (Simplified/Traditional Chinese, English, Spanish, and ~30 more) — the language you choose becomes Collie's own interface language too.
Click through. When it finishes, open Collie from the Start menu or the desktop icon.
On first launch, pick a brain: an existing Claude, Codex, or Grok subscription connects in one click, or paste an API key. That's it.

Everything Collie needs ships inside the installer — an embeddable Python, the semantic-memory engine, the desktop window (WebView2), and the browser extension. It installs per-user (no admin prompt) and cleanly uninstalls from Add or remove programs.

Optional extras during setup

Two checkboxes let you turn on the live star-map wallpaper and the real-browser bridge at logon. Both are off by default and covered in The desktop app.

Developers — pip / uv¶

The core is stdlib-only, so the base install is tiny.

# from source (PyPI publish is planned):
git clone https://github.com/colliehq/collie && cd collie
pip install -e ".[local,dev]"

Then let Collie finish provisioning itself:

collie setup     # installs optional deps, pre-downloads the memory model, picks a provider
collie           # opens the terminal chat

Optional extras¶

Extra	What it adds
`local`	Semantic memory — granite-107m via onnxruntime (~55 MB, multilingual). What `collie setup` installs.
`tui`	The rich terminal chat (`collie tui`).
`browser`	Playwright — only for `collie browser-bridge --browser` (a managed Chromium with the extension preloaded, for CI / no-login use). The everyday real-browser control uses your own Chrome + the extension and needs no extra.
`search`	Keyless web search (DuckDuckGo).
`acp`	Editor integration over the Agent Client Protocol (`collie acp`).
`fastembed`	Opt-in jina-v3 and other fastembed models.

pip install "collie-harness[local,tui,search]"

Semantic memory is optional

Without the local extra, memory runs on BM25 keyword recall — it works out of the box, it just isn't semantic. Collie never silently falls back to a low-quality embedder.

Verify it works¶

collie selftest      # $0 deterministic end-to-end: mock model, real tools + memory + dashboard

If that prints a green verified run, you're set → Quickstart.