Skip to content

Quickstart

This assumes Collie is installed. Every example runs locally.

Zero-config

collie                 # terminal chat (TUI); the first run picks a provider
collie web             # browser GUI — chat, the live verification gate, diffs, the star-map
collie app             # the native desktop window (Windows)

Your first verified fix

Point Collie at a real bug with a cheap model:

DEEPSEEK_API_KEY=... collie -p "fix the off-by-one in utils/timeparse.py"

Watch the verification gate: it writes a reproduction that fails, makes the smallest edit, and re-runs it. The run ends verified ✓, not "done."

Fully local, no key

collie run "summarize app.py" --provider ollama --model qwen2.5-coder:7b

Machine-readable output

collie run "fix the bug" --json          # final result object (tokens, cost, verified)
collie run "fix the bug" --stream-json   # live NDJSON: tool · edit · repro-gate · receipt

Autonomy that ends on green

# iterate toward a goal; STOP the first turn a real executed check passes
collie loop --goal "get the suite passing" --until "pytest -q" --max 8

The loop terminates on a real executed check (--until exits 0), not on the model announcing it's finished — the verification gate, one level up.

Best-of-N by what actually passes

# run 3 isolated attempts, keep only what passes --check; apply nothing if none pass
collie pack "fix the failing test" -n 3 --check "pytest -q" --apply

Warm a repo (optional)

collie init            # pre-download the memory model for this repo
collie init --rules    # additionally have the model write an AGENTS.md

collie init is optional — the first question pays the indexing cost otherwise. Collie reads AGENTS.md / CLAUDE.md as project rules on every run.

Where next