Blame for MCP_SERVER.md

# Options Scanner — MCP Server

[![Built with ML Intern](https://huggingface.co/datasets/huggingface/badges/resolve/main/built-with-ml-intern-sm.svg)](https://hf.co/chat?mode=ml-intern)

The [`options_scanner`](https://huggingface.co/datasets/Reckless88/options-scanner) engine
wrapped as a **modular MCP server**: any MCP client (Claude Desktop, Claude Code,
Hermes, etc.) can scan markets, build multi-leg options structures, and manage
positions with rolling mechanics through natural language.

**Read-only by design.** The server reads market data and computes structures.
It never places orders, touches a broker, or writes anywhere on disk.

---

## Tools (7)

| Tool | What it does |
|---|---|
| `scan_universe` | Run the full multi-strategy scanner over the auto universe (large caps + high-IV names + ETFs) or a custom symbol list. Returns ranked candidates with credit/debit, max profit/loss, POP, breakevens, legs. |
| `scan_symbol` | Build every eligible strategy on one ticker with full risk reports and gate reasons. |
| `build_strategy` | Build one named family on a symbol: `vertical`, `calendar`, `diagonal`, `butterfly`, `iron_condor`, `ratio_backspread`, `jade_lizard`, `double_diagonal`, `short_straddle`, `short_strangle`. |
| `chain_summary` | Quick read on a ticker: spot, 30-day ATM IV, front-expiry liquidity score, next earnings, quote count. |
| `check_position` | Position management with rolling mechanics: take-profit (50% of credit), stop (2× credit), roll out / out-and-down / out-and-up at 21 DTE or when a short strike's delta hits 0.40. |
| `list_strategies` | All 10 strategy families + default risk parameters and roll rules. |
| `list_watchlists` | Built-in watchlist pools and the auto universe. |

Every tool takes an optional `provider` argument (`tradier` | `alpaca` | `polygon` | `yfinance`); empty = auto-pick from environment keys.

## Install

```bash
git clone https://huggingface.co/datasets/Reckless88/options-scanner
cd options-scanner
pip install "mcp>=1.0" yfinance          # yfinance only needed for the fallback provider
export TRADIER_API_KEY=...               # optional; see providers below
python mcp_server.py                     # stdio transport — this is what clients launch
```

**Data providers** (auto-selected in this order):

| Env var(s) | Feed |
|---|---|
| `TRADIER_API_KEY` | Tradier real-time (verified: full chains, OI; greeks computed from quotes) |
| `ALPACA_KEY_ID` + `ALPACA_SECRET_KEY` | Alpaca real-time IEX (thin quotes, no OI) |
| `POLYGON_API_KEY` | Polygon real-time snapshot |
| none | Yahoo Finance, ~15-min delayed |

No keys? The server still works on the delayed yfinance feed.

---

## Import into Claude Desktop

1. Install the repo somewhere stable, e.g. `~/tools/options-scanner`, and put your key in the file (see security note).
2. Edit the config file:
   - **macOS:** `~/Library/Application Support/Claude/claude_desktop_config.json`
   - **Windows:** `%APPDATA%\Claude\claude_desktop_config.json`
3. Add:

```json
{
  "mcpServers": {
    "options-scanner": {
      "command": "python3",
      "args": ["/Users/YOU/tools/options-scanner/mcp_server.py"],
      "env": { "TRADIER_API_KEY": "your-key-here" }
    }
  }
}
```

4. Restart Claude Desktop. A tools icon appears; ask *"Scan SPY, QQQ and NVDA for defined-risk strategies"* or *"Build an iron condor on COIN"*.

**Claude Code** (CLI) — one command instead of the JSON:

```bash
claude mcp add options-scanner \
  --env TRADIER_API_KEY=your-key-here \
  -- python3 /Users/YOU/tools/options-scanner/mcp_server.py
```

---

## Import into Hermes

**Option A — one command:**

```bash
hermes mcp add options-scanner \
  --command python3 \
  --args /path/to/options-scanner/mcp_server.py \
  --env TRADIER_API_KEY=your-key-here
```

**Option B — edit `~/.hermes/config.yaml`:**

```yaml
mcp_servers:
  options-scanner:
    command: "python3"
    args: ["/path/to/options-scanner/mcp_server.py"]
    env:
      TRADIER_API_KEY: "your-key-here"
    enabled: true
    timeout: 120
```

Then start `hermes chat`. Hermes discovers the tools with an `mcp_options-scanner_*` prefix and you ask the same natural-language questions.

---

## Security notes

- API keys go in **client config env blocks**, never in the source. The repo contains no keys.
- The server is read-only: it has no broker credentials, no order paths, no disk writes.
- `allow_undefined_risk=True` unblocks naked straddles/strangles in *analysis only* — they are always flagged `UNDEFINED RISK` in the output.

---

## Evaluation of the underlying engine (honest assessment)

**Verified working** (tested against live Tradier and Alpaca feeds, Oct 27 2026):
- Full scan of 40 underlyings in ~40 s → 84 candidates across 13 strategy families, 0 feed errors.
- Payoff-grid risk metrics with back-month legs priced via Black-Scholes; max loss floored at the debit paid.
- POP by numeric integration of each structure's actual payoff over the lognormal distribution (catches debit-spread downside breakevens the naive short-strike method misses).
- Gates: per-leg liquidity, bid-ask vs mid, finite max loss, buying-power cap, earnings-window avoidance, strike proximity (refuses to fake structures on sparse feeds).
- MCP protocol end-to-end with the official `mcp` SDK 2.3.0: 7 tools discovered and called over stdio.

**Known limitations — read before trading anything it prints:**

1. **Pricing model is approximate.** Back-month legs are valued with flat entry IV — no vol skew, no term structure, no IV crush modeling. Calendar/diagonal max-loss figures are lower bounds on reality, not guarantees.
2. **POP is model POP.** Lognormal with ATM IV. Fat tails, realized-vs-implied gaps, and event risk are not in it.
3. **No backtest.** The 0–100 composite score (expectancy 45% / POP 25% / credit efficiency 20% / liquidity 10%) is a heuristic ranking, *not* validated expected performance.
4. **Fills at mid, zero slippage.** Real 4-leg structures fill worse than mid, especially outside regular hours.
5. **No early-assignment modeling** (ITM short puts/calls near ex-dividend dates) and no pin-risk discussion for short strikes at expiry.
6. **No persistence or scheduling.** Positions are passed per call; the server holds no state and runs nothing on a timer — automation loops live in the client or a cron job.
7. **Feed-dependent coverage.** Tradier's free tier returned no greeks (computed per-quote instead); Alpaca IEX lacks open interest and has thin quotes on some names (guards drop them). Earnings dates come from a yfinance fallback that intermittently 404s — the earnings gate can be blind.
8. **Curated universe, not all of Nasdaq/NYSE.** The auto pool is ~40 liquid names ranked by outcome; it is not a full-market sweep.

**Bottom line:** a solid idea-generation and risk-sizing tool for defined-risk options structures, with honest gating and verified real-time data plumbing. Treat every number as an estimate to be checked in your broker before ordering.