Hypothesis → review through a different provider → bounded simulation → real-result diagnosis → complementary signals → pre-submission acceptance. You control models, budgets, evidence and the queue — no chat window needs to stay online.
# pick one: pipx / Homebrew / DMG $ pipx install wq-pilot $ brew tap Octo-o-o-o/autowq && brew install wq-pilot # any directory becomes your workspace $ mkdir ~/autowq && cd ~/autowq $ wq onboard
Zero third-party dependencies, pure Python 3.11+ standard library. Four install paths; macOS users can add a menu-bar app.
One command, cross-platform: pipx install wq-pilot, then wq onboard in any empty directory.
brew tap Octo-o-o-o/autowq && brew install wq-pilot; menu-bar app: brew install --cask Octo-o-o-o/autowq/worldquant.
Download the DMG from Releases, drag to Applications and double-click: the first-launch wizard picks a workspace, prepares the runtime and starts onboarding, then lives in your menu bar.
git clone the repo, then ./wq onboard; tests: PYTHONPATH=src python3 -m unittest discover -s tests.
A gated research pipeline: every step is auditable, every budget capped, every submission verified.
The research-role model proposes candidate alphas from field evidence and full-history reviews; each task freezes its own routing snapshot.
The review role must run on a different provider, so one model's blind spots never confirm themselves.
Real simulations, GET polling and result accounting; uncertain POSTs are never blindly retried.
Failure-reason feedback, time-segment evidence and daily PnL correlations flow back into the next cycle.
Bounded combination experiments with preregistered decay/neutralization variants, every result recorded.
Per-candidate submission: research review, fresh platform checks, a single POST and verified acceptance status.
Built for long unattended runs: budgets, authorization, recovery and dedup are defaults, not options.
Grok Build, Devin, Cursor, ZCode, Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Qwen Code, OpenCode, plus standard OpenAI / Anthropic APIs.
sandbox-exec confines model processes on macOS; Docker images isolate them on Linux/WSL2; credentials live in a 0700 private directory.
Single concurrency, per-provider budgets, authorization expiry, rate-limit handling; unknown costs are never booked as zero.
A local SQLite ledger tracks every task; runs resume from the ledger and UNKNOWN requests require human reconciliation.
Cycle history, submitted alphas, routing presets, provider switches, intervals and system notifications — all from the menu bar.
Task progress with token/cost provenance; export summary/tasks/results (JSON/CSV) without cookies, passwords or task inputs.
Run inside your workspace after install; every destructive capability starts disabled.
$ wq onboard --list # detect host CLIs only; no writes $ wq onboard # first-run wizard: providers, models, roles, login $ wq doctor --fix-private # environment checks, private-dir repair $ wq autopilot status # automation status $ wq tasks # task queue $ wq brain fields --datasets # read-only data-field catalog snapshot $ wq pause --reason "manual" # pause queue and local model calls
| Platform | Install | Process isolation |
|---|---|---|
| macOS | pipx / Homebrew / DMG | sandbox-exec |
| Linux | pipx / source | Docker images |
| Windows | WSL2 (same as Linux) | Docker images |
No. autowq is an independent open-source project (Apache-2.0), not affiliated with WorldQuant. WorldQuant/BRAIN trademarks, platform and data belong to their owners.
autowq itself is free and open source. It orchestrates your own model CLI subscriptions or API keys and your own BRAIN account — those costs are between you and the vendors. Onboarding never starts paid inference.
No. Configuration, cookies, real alphas, databases, model outputs and logs stay local; exports exclude cookies, passwords, task inputs and evidence paths. Model requests go only to providers you explicitly select.
Not natively. Use the WSL2 + Docker path; macOS and Linux are supported natively.
No. Passing a screen, an accepted submission, consultant eligibility and actual payment are separate states. Research automation does not grant submission privileges; submission has its own acceptance gates.
A model reviewing its own hypotheses shares its own blind spots. autowq requires research and review roles to come from different providers for continuous automation, and the wizard enforces this at configuration time.