Collie

A coding agent that lives on your computer — and can actually run it. Local and private, Collie reaches your real environment — your logged-in browser, your desktop, your screen, your files — and proves its work by running it.

- :material-rocket-launch: **[Install](install.md)** — one-click Windows installer, or `pip` from a release - :material-play: **[Quickstart](quickstart.md)** — your first verified fix in a minute - :material-brain: **[Providers](providers.md)** — connect Claude, Codex, Gemini, a local model, or an API key - :material-console: **[CLI reference](cli.md)** — every `collie` command - :material-monitor: **[The desktop app](desktop.md)** — native window, live star-map, real-browser bridge

Why it's different

Most coding agents live in a cloud tab or an editor pane and can only touch the files you hand them. Collie runs on your machine, so it works the way you do: it drives your real logged-in browser, arranges your desktop, records your screen, takes tasks from your phone, and edits your code — all locally, all under your control.

And the range is the proof. The desktop console, the browser control, the screen recorder, the phone remote — Collie's coding agent built all of it. A harness strong enough to ship its own desktop app and iOS companion is strong enough for your bug.

It proves its work: the verification gate

When Collie fixes something it writes a reproduction that must fail on the broken code, makes the smallest edit that flips it, and re-runs the assertion. A run isn't "done" — it's verified ✓.

  repro    wrote repro.py · assert parse_duration("1h30m") == 5400
           ✗ FAILING  › got 1800, want 5400              ← gate armed

  edit     utils/timeparse.py  ································· +1 −1
           43 │- total = SECONDS[unit] * int(val)
           43 │+ total += SECONDS[unit] * int(val)

  verify   python repro.py
           ✓ PASSING  › parse_duration("1h30m") == 5400   ← gate green

  ✓ verified in 12.8s · Δ +1 −1 · 3,410 tok · $0.006

Other agents check "did the test not error." Collie's gate is stronger: the reproduction carries an assert actual == expected derived from the issue, so a plausible-but-wrong edit fails loudly and drives another repair round.

What makes it lean

• Executed verification, not self-reported success — the one change that measurably moved Collie's SWE-bench resolve rate.
• Zero third-party dependencies in the core. mock and ollama run with no key; memory works out of the box on BM25 keyword recall, and upgrades to real semantic recall when you install the optional local embedder.
• Model-agnostic. Claude, GPT/Codex, Gemini, DeepSeek, Qwen, a local Ollama model — same harness, so any delta is the harness, not the model.
• Runs locally, no telemetry, no account.

Where to go next

New here and just want it running? → Install. Already installed? → Quickstart.