APEX-1

An autonomous options-trading agent where Claude is the decision-maker.

A scheduler wakes Claude on a cron schedule through the trading day. Each session Claude reads market data, computes its own signal, reasons through risk and signal gates, places call-option orders through Alpaca, manages positions, and reflects on performance — with no human in the loop. The only hard human control is a single environment flag that gates live trading.

Status: paper-validation. The system trades paper only and does not scale or enable live capital until ≥ 20 closed trades show positive measured expectancy.

Current architecture baseline: On 2026-06-10, commit d05363d rebuilt APEX-1 around an owned GEX + momentum signal and removed Unusual Whales and yfinance from the active system.

---

What it is

APEX-1 is built on three ideas:

1. Claude is the brain, not a notifier. A thin Python loop (agent.py) hands Claude a session prompt and a toolbox, then executes the tool calls Claude makes until it decides the session is done.
2. The signal is fully owned. No external signal vendors. APEX-1 computes a dealer-gamma (GEX) map plus momentum entirely from Alpaca data.
3. Survival first. Hard risk limits (position size, count, drawdown, calls-only, DTE, forced flat by 3:45pm) are enforced in code — the model cannot override them.

Architecture

Scheduler (scheduler.py)  →  wakes  →  Agent Loop (agent.py)  →  calls  →  Tools (tools/)
                                              ↕
                                    Anthropic Messages API
                                    (Sonnet for reasoning, Haiku for monitoring)

Sessions (ET, Mon–Fri): signal generation 8:15 · premarket 8:30 · trading 9:45 & 14:00 · monitor every 5 min · force-close 15:45 · postmarket 16:30 · evolution 17:00.

The signal: GEX + momentum

A slow structural map and a fast trigger.

• Structural (leading) — tools/gex.py. Pulls calls + puts with open interest from Alpaca, computes gamma locally (Black-Scholes), and aggregates to net GEX, the zero-gamma flip, and the call/put walls. Positive net GEX = dealers suppress volatility (price pins / mean-reverts); negative = dealers amplify it (price trends / breaks out).
• Trigger (confirming) — tools/market.py. Opening-range breakout, volume ratio, and 5m/1h momentum confirm the entry. Momentum never enters alone.
• Setups (tools/signal.py, calls only): negative_gamma_breakout (net GEX < 0, spot above the flip, momentum up) and positive_gamma_drift (net GEX > 0, spot below the call wall with room, momentum up).
• Universe: liquid index ETFs (SPY, QQQ, IWM) — deep gamma, tight spreads, no earnings risk.

Safety model

Enforced in the tool layer (tools/alpaca.py), not just in prompts:

• Forced paper mode unless LIVE_TRADING_ENABLED=true.
• Max premium $40 live / $160 paper, calls only, DTE ≥ 3, max 3 open positions, 8% daily drawdown halt.
• Force-close at 3:45pm ET — no overnight holds.
• Compounding-aware sizing (suggest_position_size) risks 1% of equity per trade, always clamped under the hard premium ceiling.
• A two-tier review "council": inline Signal/Risk gates (zero API cost) plus independent Trade-Reviewer and Evolution-Validator Claude calls whose BLOCK verdicts are binding.

External dependencies

Service	Role	Notes
Anthropic API	Reasoning brain	Sonnet (reasoning) / Haiku (monitoring)
Alpaca	Broker + all market data	Execution, stock bars, option contracts/OI, real-time OPRA quotes (Algo Trader Plus)
Telegram	Audit notifications	Not an approval gate

No Unusual Whales, no yfinance. Those external vendors were removed from the active system on 2026-06-10 in commit d05363d (Rebuild on owned GEX + momentum signal; remove external vendors). VIX is not offered by Alpaca, so the volatility regime gauge is SPY ~30-day ATM implied vol computed locally (see docs/CHANGELOG.md).

Repository layout

agent.py            Agentic loop (Anthropic Messages API)
scheduler.py        APScheduler cron triggers (ET)
config.py           Constants: safety limits, sizing, universe, regime thresholds
tools/
  gex.py            Dealer-gamma map
  signal.py         GEX + momentum signal engine
  market.py         Regime, momentum, earnings (Alpaca)
  alpaca.py         Execution, chains, real-time quotes, sizing
  blackscholes.py   Shared delta/gamma/IV math
  memory.py         State files (file-locked)
  trades_db.py      Trade journal queries
  telegram.py       Audit messaging
council/            Two-tier review governance
prompts/            System + per-session prompt templates
db/schema.py        SQLite schema
tests/              pytest suite
docs/
  CHANGELOG.md      Change history
  plans/            Design/rearchitecture plans
  archive/          Superseded specs + historical audit
CLAUDE.md           Developer/agent guide

Setup

pip install -r requirements.txt

Create a .env in the project root (never commit it — it is gitignored):

ANTHROPIC_API_KEY=...
ALPACA_API_KEY=...
ALPACA_SECRET_KEY=...
ALPACA_BASE_URL=https://paper-api.alpaca.markets
TELEGRAM_BOT_TOKEN=...
TELEGRAM_CHAT_ID=...
LIVE_TRADING_ENABLED=false

Running

python db/schema.py                       # initialize the database
python agent.py --session trading --dry-run   # mocks execution, real data
python scheduler.py                       # run all sessions on cron
python -m pytest tests/ -v                # test suite
python -c "from tools.gex import compute_gex; print(compute_gex('SPY'))"

Docs

• docs/CHANGELOG.md — change history
• docs/plans/ — design and rearchitecture plans
• CLAUDE.md — developer/agent guide

Disclaimer

APEX-1 trades real financial instruments and can lose money. It is a research system provided as-is, with no warranty. Nothing here is financial advice. Run it on paper and validate a positive edge before risking live capital.