QQuantprims

AI calc infrastructure

Don't trust AI-written backtest.Use verified primitives.

Deterministic finance primitives for Claude, Cursor, Codex, ChatGPT — backed by 9 months of dogfooded FX backtest engine.

MCP server is shipping first (Claude Desktop, Cursor, Codex, ChatGPT Desktop). REST API for browser / curl / any framework is coming soon. 7-day trial, then ¥980/month.

Request early accessPreview docs (coming soon)

Same numbers every time. 100× faster than AI rewriting. Audit-ready.

claude-desktop

You: Sharpe ratio of EURJPY with RSI(14) strategy?

Claude: calling fx_backtest.run

{
  "strategy": "rsi_oversold",
  "symbol":   "EURJPY",
  "params":   { "rsi_period": 14 }
}

✓ verified, deterministic, 187 ms

Claude:

Sharpe:      1.42
Max DD:      -8.3%
Win rate:    54.1%
Profit factor: 1.78

How it works

Stop letting AI reinvent your Sharpe ratio.3 steps to deterministic finance.

I built this because I was tired of Claude reinventing Sharpe every chat session. Now my AI agent calls verified primitives instead.

  1. Request early access

    Email me from the dashboard page. The founder replies personally and issues a trial license key by hand while the self-serve checkout dashboard ships.

  2. Wire it into your AI

    Add the MCP server to Claude Desktop / Cursor / Codex config. Public REST API for ChatGPT and any framework is coming soon.

  3. Ask. AI does the rest.

    Tell your agent what to test. It fetches data, calls our verified primitives, and returns deterministic results.

Calc primitives

8 primitives. Same numbers every time.

Verified engine, not AI-generated code. 9 months of dogfooded development. 303 unit tests + multi-round adversarial review.

  • run_backtest

    Execute strategy across OHLC, return trades + equity curve.

  • calc_sharpe

    Risk-adjusted return — annualized, deterministic, audit-ready.

  • calc_maxdd

    Peak-to-trough drawdown + duration + recovery time.

  • run_walkforward

    Anchored or rolling WFO with purged splits — no lookahead.

  • run_montecarlo

    Bootstrap returns, get distribution of outcomes + percentiles.

  • optimize_grid

    Brute-force parameter sweep with explicit budget cap.

  • optimize_genetic

    Evolutionary search over high-dim parameter space.

  • compare_strategies

    Side-by-side stats across N strategies, single call.

Two transports, one engine

MCP today, REST coming soon.

Call our primitives from Claude Desktop, Cursor, Codex, or any MCP-compatible IDE today. The public REST API (for ChatGPT Custom GPT Actions, plain curl, and any framework) is in active development. Both transports share the same shared/ core.

MCP server

stdio · live in trial

Plug into Claude Desktop, Cursor, Codex CLI, ChatGPT Desktop, or any MCP-compatible IDE.

{
  "mcpServers": {
    "quantprims": {
      "command": "npx",
      "args": ["@quantprims/mcp"],
      "env": { "LICENSE_KEY": "fxb-…" }
    }
  }
}

REST API

HTTPS + JSON · coming soon

Will let any framework, language, or web UI call the same primitives. In active development — endpoints, OpenAPI spec, and SDK samples publish before launch.

# Coming soon — endpoints and auth will be
# published in the API docs before launch.
# Example shape (subject to change):
#   POST /v1/calc_sharpe
#   Authorization: Bearer <license-key>
#   { "returns": [...] }
Both backed by the same shared/ core — 9 months of dogfooded TypeScript primitives.

Why trust this

AI is fast. AI is wrong sometimes.Wolfram-grade primitives fix that.

Same story as Wolfram Alpha for math — when correctness matters, you call a verified engine, not let the LLM reinvent it.

  • Deterministic

    Same input → same output, every time. No AI hallucination drift between sessions.

  • 100× faster iteration

    Verified primitives return in ~0.1s. AI re-implementing Sharpe each time takes 5-20s.

  • Audit-ready

    9 months of TypeScript primitives. 303 unit tests. Multi-round adversarial review log.

  • Safe boundary

    Schema-constrained inputs + read-only data. No prompt injection, no arbitrary execution.

Pricing

Founder dogfood pricing.

Pricing is set. Self-serve Stripe checkout is in active development; until it ships, early access is by manual founder review.

Free trial

7 days, no card required (after launch)

¥0for 7 days
  • All 8 calc primitives
  • MCP server (live in trial) + REST API (coming soon)
  • Up to 2 devices per license
  • Audit log + reproducibility
  • Email support
  • Cache-friendly tool descriptions (= compatible with Anthropic Prompt Caching = up to ~90% input-cost reduction on the cached prefix when AI agents call us repeatedly)
Request early access
Recommended

Pro

Founder dogfood pricing (planned)

¥980/ month
  • All 8 calc primitives
  • MCP server (live in trial) + REST API (coming soon)
  • Up to 2 devices per license
  • Audit log + reproducibility
  • Email support
  • Cache-friendly tool descriptions (= compatible with Anthropic Prompt Caching = up to ~90% input-cost reduction on the cached prefix when AI agents call us repeatedly)
Request early access

Note: during early access, this service is intended for personal/evaluation use only. Commercial and business workloads will be addressed once the lawyer-signed opinion on Japanese investment advisory law non-applicability is published and the Terms of Service explicitly define the commercial scope. Lawyer opinion under Article 28, Section 3 of the Financial Instruments and Exchange Act is in progress.

FAQ

Common questions

Can I use this from ChatGPT, not just Claude?

Eventually yes via the REST API, which is in active development and not yet live. The MCP transport is live in trial today and covers Claude Desktop / Cursor / Codex CLI / IDE clients. ChatGPT Custom GPT Actions and curl access ship when the REST API publishes.

How do I wire it into Claude Desktop?

Add a single block to claude_desktop_config.json with the MCP server command + your LICENSE_KEY env var. The full setup snippet is sent manually with the early-access license key (the public docs page is still being prepared). Setup takes under 2 minutes once you have the snippet.

Can I use this commercially?

Not yet. During early access, the service is intended for personal/evaluation use only. Commercial and business workloads will be addressed once the lawyer-signed opinion on Japanese investment advisory law (= Article 28 Section 3, Financial Instruments and Exchange Act) non-applicability is published and the Terms of Service explicitly define the commercial scope.

Isn't this giving investment advice?

No. The engine returns numbers (= Sharpe / max drawdown / win rate / etc.) — never recommendations, evaluations, or buy/sell signals. The 'number-only' design choice is deliberate: outputs are statistics, not advice. Interpretation is on you (or your AI agent). A lawyer-signed opinion on Japanese investment advisory law (= Article 28 Section 3, Financial Instruments and Exchange Act) non-applicability is in progress; will be published once obtained.

Where does the price data come from?

We don't supply price data. You (or your AI agent) pass OHLC arrays as input — sourced from your broker export, paid data API, or your own collection. Same structure as Wolfram Alpha for math: bring your numbers, we calculate. Data source compliance is your responsibility.

What's covered by the license?

Early-access licenses cover all 8 calc primitives via the MCP server, up to 2 devices per license, with full audit log + reproducibility. Self-serve trial/paid checkout, Public REST API access, and any post-launch entitlements (including refund/credit terms) will be defined in the Terms before they ship.