Cents

Research tool, not an investing tool

Cents is an open-source experiment in agent-orchestrated investment research. It is not an investment adviser, broker, or recommendation engine, and nothing it emits is investment advice. Real-money trading is technically possible but explicitly out of scope — the autonomous loop is hard-coded to paper, and the controls a real trading system would require (vol-scaled sizing, drawdown gates, slippage and borrow modeling, reconciliation, audit recordkeeping) are intentionally absent. Read the full scope statement before going further.

Cents is a CLI sandbox for treating investing as a hypothesis-testing exercise — write down a thesis, declare which regime variables it depends on, and let the system flag when one of those dependencies shifts. Pluggable screeners discover candidate universes, specialised research agents gather evidence, and an autonomous factory loop walks the cycle in paper mode only — producing a labeled outcomes dataset you can study.

git clone https://github.com/wolfbane/cents.git
cd cents && pip install .

Demo recording coming soon.

The asciinema cast at /demo.cast is a placeholder. Once recorded, it will play here in-browser.

Thesis-driven

Every position starts as a written hypothesis with a valuation, horizon, target, and stop. Research is evaluated against the thesis, not in the abstract.

Regime-aware

Each thesis declares the policy / macro variables it depends on (tariffs.china, fed_policy, ai_capex…). An EventAgent ingests Federal Register policy events; when an event hits one of your tags, you get a premise-invalidation alert — before the market fully prices it.

Multi-agent

Eight specialised agents — fundamentals, technical, macro, sentiment, moat, insider, event, plus an orchestrator — pull from FMP, Alpaca, FRED, NewsAPI, and the Federal Register; each contributes evidence and a conviction delta.

Paper-only factory

cents factory run walks a universe of symbols and opens paper theses where the orchestrator clears your entry threshold (with neutral twins in paired mode), closing on target / stop / horizon / premise-invalidation. Hard-coded to paper — real trading is out of scope.

Pluggable screeners

Five built-in discovery strategies — value, growth, momentum, mean-reversion, insider-cluster — that turn a parent universe into a candidate list the factory walks. Every thesis records which screener produced it, so cents factory analyze --by discovery surfaces which discovery strategy wins in which regime.

Accuracy-tracking

Outcomes are recorded with a closure reason — correct / incorrect / unclear / invalidated / preempted — plus the regime snapshot captured at thesis birth, so you can stratify later instead of averaging across regimes that have nothing to do with each other.

Cost-visible

Every Anthropic call is recorded to an llm_usage table; cents usage summary reports input/output/cache tokens and estimated cost by agent, model, day, or operation. Nothing about the cost of autonomy is hidden.

Get started

Scope

What cents is, what it is not, and why real-money trading is out of scope. Read this first. Read the scope →

Quickstart

A 5-step walkthrough from zero to your first paper scan. Read the quickstart →

How it thinks

The operating principles behind regime-awareness, premise tags, and the factory’s labeled-outcomes goal. Read the principles →

Star on GitHub

Source, issues, and discussions live at github.com/wolfbane/cents.