Collie¶
Your private, local-first personal intelligence system. Collie learns the workflows you choose to share, coordinates work across your computer and devices, asks when authority is needed, and keeps source and verification evidence attached to the result.
Why it's different¶
Memory lives inside the workflow. Optional outside-AI learning turns content-free local activity and separately connected sources into reviewable habits, suggested routines, and evidence-backed reminders. Raw observations stay on the device, and a learned routine begins with zero authority.
Mission-first, not chat-first. Desktop, terminal, IDE, phone, and browser enter the same durable Mission history. Work can survive waits, retries, restarts, and handoffs; decisions that need your authority enter one global Needs You inbox.
One Collie, a Pack underneath. The user sees a stable Collie identity. Models, specialist workers, skills, app connections, and devices remain replaceable execution resources behind it.
Local reach with an explicit leash. Collie can use your real logged-in browser, desktop, screen, files, and code. Sensitive actions are bounded by permissions and budgets; task context goes only to the model provider and services you connect.
Evidence you can inspect: the verification gate¶
For code changes, Collie can record a failing reproduction, edit the implementation, and execute the named assertion again. Required verification blocks completion until the selected post-edit check passes. The result is evidence for that check's stated scope—not proof of every possible property.
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
The receipt should say what ran, whether it passed, what it covers, and what remains unverified. A green gate without that scope is only a confidence symbol; it is not an accountable completion rule.
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.
mockandollamarun 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 with no telemetry or account; Online is separately opt-in.
Where to go next¶
New here and just want it running? → Install. Already installed? → Quickstart.