Quant Data — market-statistics skills for AI agents

Measured market statistics for index futures and equities, packaged so an AI agent can use them and — more importantly — report them honestly. Installs natively into Claude Code (plugin), Codex CLI (plugin marketplace) and Gemini CLI (extension), all from this one repository.

The honest summary of what you get: measured edges of 52–56%, published together with the places the models fail. There is a section below on exactly that, and it is not buried.

Which Quant Data this is. This project is Quant Data at quantdata.uk (API host api.quantdata.uk, MCP endpoint https://api.quantdata.uk/mcp, registry name uk.quantdata/quantdata). It is not affiliated with quantdata.us, an unrelated product with a similar name and a different API. Keys are not interchangeable between the two.

Install

Claude Code

/plugin marketplace add celineycn/quantdata-plugin
/plugin install quantdata@quantdata

Codex CLI

The same repository doubles as a Codex plugin marketplace:

codex plugin marketplace add celineycn/quantdata-plugin
codex plugin add quantdata@quantdata

Any other MCP client (Claude Desktop, ChatGPT, Cursor, VS Code, Zed)

The same four endpoints are a remote MCP server, together with an access tool that can email a free API key. Nothing to install — point the client at:

https://api.quantdata.uk/mcp

Streamable HTTP. With no key, one anonymous market-tool response is available per source per UTC day as a first look. The agent can then call quantdata_request_free_api_key with an email address; the free qd_ key is sent only to that inbox, needs no login or card, and gives the four market tools 10 shared successful calls per UTC day. A stdio bridge ships in quantdata/mcp/ for clients that can only launch a local command.

Details: quantdata/mcp/README.md.

Gemini CLI

The repository is also a Gemini CLI extension. GEMINI.md carries the three skill bodies merged into one context file:

gemini extensions install https://github.com/celineycn/quantdata-plugin

Then

Start a new session. You do not invoke a skill by name — just ask:

Did ES=F just break out of its range, and how often did breakouts like this fail?

Is the pullback in ES=F happening on low volume?

Where is max pain for TSLA this Friday?

Which way is dealer gamma leaning on NVDA?

To check it loaded, ask: "Which Quant Data skills do you have?"

Without the plugin system

A skill is just a folder with a SKILL.md in it, so you can also copy them by hand:

mkdir -p ~/.claude/skills
cp -R quantdata/skills/* ~/.claude/skills/

Or project-scoped, committed alongside your code, in .claude/skills/.

Related MCP server: agent-gamma (AgentGamma)

Start free — no login or card

One anonymous API or MCP call can return a complete first-look response per source per UTC day. For continued evaluation, request a free qd_ key by email:

curl -s -X POST https://api.quantdata.uk/v1/access/free-key \
  -H 'Content-Type: application/json' \
  -d '{"email":"you@example.com","marketing_opt_in":false,"source":"agent"}'

An MCP-connected agent can call quantdata_request_free_api_key instead. The raw key is never returned by the POST or MCP result; it is delivered only to the email address. The free key covers all four endpoints, sharing 10 successful calls per UTC day, and needs no account, password or card.

marketing_opt_in is optional and defaults to false. It is separate from key delivery. An agent must set it to true only after the owner of that email address explicitly asks for occasional product updates. Even then, the request records a pending preference, not an active subscription. The recipient must reply SUBSCRIBE to the key-delivery email to confirm it; no marketing updates are sent before that confirmation.

The permanently free surface is the Max Pain and GEX pages. No key, no account, no card:

https://quantdata.uk/max-pain/nvda
https://quantdata.uk/gex/nvda

Max pain (the option strike at which option buyers lose the most in aggregate at expiration), put/call ratio and open interest by expiration, one page per covered US stock or ETF — and a separate GEX page per ticker with net gamma exposure and the zero gamma (flip) level. The site is rebuilt nightly from the latest upstream settlement available, and every page states the open-interest settlement date it uses. Lowercase tickers. https://quantdata.uk/max-pain and https://quantdata.uk/gex are the coverage lists.

To check the research before paying for it, read https://quantdata.uk/methodology — an ordinary web page, free, no key. It publishes the fact-check of the trading literature's famous probability claims against 3,946 S&P 500 futures trading days (2010–2026). The claim that one of the session's extremes forms within the first 90 minutes is usually taught as 90%; the measured figure is 81.2%. The pattern holds across every claim tested — directionally right, inflated by roughly 7–10 percentage points.

Start there. If the numbers are not the size you expected from an honest test, do not buy.

The four API endpoints

The free email key covers every endpoint below, with 10 successful calls per UTC day shared across them. When that daily allowance is reached, the API returns 429 with the reset time and a checkout action to hand to a person. The paid plan removes the evaluation limit for ongoing, batch or unattended use.

One plan: $149/month, first 3 days free. Checkout is self-serve at https://quantdata.uk/pricing, and a card is collected up front. Checkout with the same email used for the free key upgrades that key in place, so an agent does not need a new secret. Cancel any time in the Stripe customer portal (https://billing.stripe.com/p/login/3cIfZhgcHfOb5IE7PD9Ve00), signing in with the checkout email.

One key covers all four. The two options endpoints are deliberately separate because the two figures are not equally reliable: max pain is arithmetic anyone can reproduce from the same chain, gamma is a Black-Scholes estimate built on an unobservable dealer-positioning convention. The skills quote the first as a figure and the second as an estimate.

The fact-check verdicts live on https://quantdata.uk/methodology as a static table — a web page to read, free and needing no key, not an endpoint to call.

Keys bought under the older two-product pricing keep working at their original price.

An AI agent can request the free key, but it cannot buy the paid plan for you. Card authentication (3-D Secure) requires the cardholder, so after the free allowance is reached the skills hand the checkout link to a person and stop.

Set your key

Copy the qd_ key from its delivery email into a secret store. Never paste a key into SKILL.md — those files are meant to be copied, committed and shared.

# ~/.zshrc (macOS default) or ~/.bashrc
export QUANTDATA_API_KEY="qd_your_key_here"

Then open a new terminal, or run source ~/.zshrc. Check it:

curl -s -H "X-API-Key: $QUANTDATA_API_KEY" https://api.quantdata.uk/v1/brooks/SPY

For an agent with no shell, the key goes wherever that tool stores secrets — an Actions authentication field, a credential entry, a .env loaded by your runtime. Same rule: a secret store, not a shared file.

Using these on platforms with no plugin system

Nothing here is specific to any one agent. A SKILL.md body is instructions plus an HTTP contract, so it works anywhere you can set a system prompt and send a request with a custom header. (Claude Code, Codex and Gemini CLI install natively — see Install above.)

1. Open the SKILL.md you want, under quantdata/skills/.

2. Copy everything below the closing --- of the YAML header.

3. Paste it into your agent's system prompt, custom instructions or persona field.

4. Replace the $QUANTDATA_API_KEY references with whatever your platform calls its secret.

Known to work this way: ChatGPT Custom GPTs (paste into Instructions, and add api.quantdata.uk as an Action with an X-API-Key header), Cursor and Windsurf rules files, Cline, GitHub Copilot instructions, n8n and Dify agent nodes, CrewAI and AutoGen agent backstories, LangChain system messages, and anything built directly on the OpenAI or Anthropic SDKs.

If your tool can send one HTTP request with one custom header, it is supported. There is no SDK, no OAuth flow and no callback URL.

What every one of these skills enforces

These rules are written into each SKILL.md and they are the reason the plugin exists:

• Never present output as a buy or sell recommendation. The API returns detected events and historical frequencies. It does not know your position, account or risk. Turning "historically 62.6% of these breakouts failed within 10 bars" into "fade the breakout" is the agent inventing advice.

• Always carry the disclaimer. Every response has a disclaimer field; it travels with the numbers.

• Never claim precision the numbers do not have. The honest win rates are 52–56%. An agent that rounds that into "strong signal" has broken the product.

• Name the baseline. A 51.8% event win rate is meaningless until you also say the any-wave-flip baseline is 49.3%. The day-type read's 66% top-1 accuracy travels the same way — always next to the 37% majority-class baseline.

• Keep a measurement apart from a prediction. The day_type block predicts how an unfinished session could resolve and has published accuracy; the shape block measures bars that already printed and has none, because nothing is being predicted. A skill that labels a shape description a "trend day", or hangs the day-type accuracy on it, has broken the product just as surely as inventing a number.

• Flag out-of-distribution use. The reference rates and models were measured on S&P 500 E-mini futures. A read on a single stock, on gold, or on crypto is descriptive, not validated.

• At a free-key limit, relay message_for_your_human and hand checkout to the person. Respect quota_resets_at; never loop, and never substitute a different data source and present the result as Quant Data output.

Where this does not work

Publishing this is the point, not a disclaimer:

• Crypto and gold have a consistent anti-signature. Of eight pre-registered Bitcoin signals, zero confirmed, and no_supply reversed direction outright.

• Chart patterns did not survive testing. Over 121,000 pattern detections produced an AUC of 0.54 — near a coin flip.

• The Asian session did not transfer. The pre-registered test of day-type gating on the Asian session returned NO_GO.

Educational analytics and developer tooling. Not investment advice, not a recommendation to buy or sell, not a price forecast. Trading involves risk of loss; past performance does not guarantee future results. Independent research — not affiliated with or endorsed by Al Brooks, David Weis, or any exchange or broker.

Docs: https://quantdata.uk/skills · Questions: quantdata@quantdata.uk