stoquant-mcp

Model Context Protocol (MCP) server for StoQuant — gives Claude (and any MCP-compatible agent) direct, read-only access to StoQuant's quant signals, ML predictions, screening, ownership, and macro data.

Free to start. A free StoQuant API key runs the core tools (quotes, Q-Score, screeners, market regime) at 100 requests/day. Power removes the cap and unlocks the premium tools (ML, alt-data, insider clustering). Get a key at https://stoquant.com/account/api-keys.

Source & listings: GitHub · npm · Official MCP Registry (io.github.ohadf2015/stoquant-mcp) · Landing page

One-command install

npx -y stoquant-mcp install

You'll be prompted for your API key. The installer writes the right claude_desktop_config.json for your OS, backs up any existing config, then prints the next step.

After install, fully quit and relaunch Claude Desktop. Type list available tools to confirm the stoquant_* tools are loaded.

Manual install (alternative)

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "stoquant": {
      "command": "npx",
      "args": ["-y", "stoquant-mcp", "serve"],
      "env": {
        "STOQUANT_API_KEY": "sk_live_…"
      }
    }
  }
}

Config locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Tools (61)

Most tools are read-only fetches. run_backtest persists a simulation run (no live or paper orders). There is no order placement, cash move, or holdings mutation. Each tool is annotated readOnlyHint so agents know it is safe to call.

Accuracy and the no-edge gate

get_accuracy_summary hits the public /public/accuracy/summary scorecard. Read excessSpreadPp (alpha spread) and dataSufficient — do not treat hitRate as skill. When dataSufficient && excessSpreadPp < 0, auto-portfolio buys are suppressed (gated:no_edge). Call this before any buy-oriented portfolio read.

Portfolio/backtest/strategy tools are evaluate/read/backtest/optimize. This server never places orders. POST /auto-portfolio/evaluate executes paper trades and is not exposed.

Two things to know

• The Q-Score (get_qscore) is the headline 0-100 composite, computed on sector-residualized returns (sector beta removed). Start here for "is this a good stock".
• The ML prediction (get_ml_prediction) returns the probability of outperforming the S&P 500 over a horizon (classification: bull|bear) — it is not a buy/hold/sell rating.

Resources

The server exposes reference docs an agent can read on demand:

• stoquant://glossary — what Q-Score, ML outperformance, regimes, and signal accuracy mean, plus unit conventions.
• stoquant://methodology — sector residualization, IC-weighting, ML health gating, and screener field semantics.

Prompts

Pre-built workflows that chain the tools:

• analyze_ticker(ticker) — full single-stock deep dive (Q-Score + ML + ownership + edges → verdict).
• find_opportunities(style) — generate and vet a shortlist for a style (value, momentum, growth, hidden gems).
• market_briefing — top-down snapshot (regime + macro + benchmarks + conviction).

Using the screener

run_screener filters on exact camelCase field names. An unknown field is treated as null server-side and silently excludes every stock, so use the documented names:

trailingPE, priceToBook, pegRatio, marketCap, debtToEquity, currentRatio, grossMargins, operatingMargins, revenueGrowth, earningsGrowth, rsi14, fiftyDayAverageChangePercent, twoHundredDayAverageChangePercent, fiftyTwoWeekHighChangePercent, shortPercentOfFloat, numberOfAnalystOpinions, sector, regularMarketPrice, regularMarketVolume, regularMarketChangePercent.

Margins/growth are decimals (0.25 = 25%); marketCap is raw USD (250000000 = 250M). Operators: gt, gte, lt, lte, eq, between (value: [low, high]), in (value: [string, …]).

The screener reads from StoQuant's cached daily fundamentals (fast, refreshed daily). Prefer run_prebuilt_screen (pre-tuned) or run_screener with universe: "sp500" and a few filters.

Example prompts

• "Get the Q-Score for NVDA and explain the top features driving the ML call."
• "Show me the top 10 conviction stocks and their current ML outperformance odds."
• "What's the current market regime, and which sectors are leading?"
• "Run the undervalued-gems prebuilt screen on the sp500 universe."
• "Screen sp500 for trailingPE under 15 and debtToEquity under 0.5."
• "Get insider activity for PLTR over the last quarter."
• "What is the live accuracy grade and excessSpreadPp? Do not buy if the tail is inverted."
• "Show auto-portfolio last eval and holdings. If gated:no_edge, do not recommend buys."
• "List backtest templates and run momentum on sp500 for 2023."

Security

• Server requires STOQUANT_API_KEY — fails fast on boot if missing or malformed.
• All requests use HTTPS. Bearer token in Authorization header. Never logged.
• Tickers are validated (^[A-Za-z0-9.\-^]{1,10}$) before any external use.
• Client-side rate limit: 200 req/min (configurable via STOQUANT_RATE_LIMIT_RPM).
• 30s per-request timeout (configurable via STOQUANT_TIMEOUT_MS).
• Server-side: API key is hashed (SHA-256) at rest, prefix-indexed for fast lookup, timing-safe-compared on verify.
• Revoke any key from https://stoquant.com/account/api-keys at any time.

Environment variables

License

MIT