Setting up your AI agent

Quant Data is built to be used through an AI agent. You ask a question in English; the agent makes the HTTP call, reads the JSON, and reports it back with the caveats attached. This page is the practical walkthrough, whatever agent you use.

On this page:

• Step 0 — read the free pages first
• Step 1 — get a key
• Step 2 — put the key where the agent can find it
• Step 3 — install a Skill, or paste the prompt (per agent)
• The prompt block
• Step 4 — ask a question
• Step 5 — read the answer critically
• Troubleshooting

Step 0 — Read the free pages first

A good part of this site needs no key, no account and no card, so you can look before deciding whether any of it is worth paying for:

• The max pain pages — static pages for 25 US stocks and 18 ETFs, rebuilt nightly. No key, no account, lowercase tickers. The quantdata-max-pain Skill works against them immediately, and an agent can keep reading them unattended forever.
• The GEX pages — estimated dealer gamma exposure for the same tickers, rebuilt on the same nightly schedule, computed by the same code the paid /v1/gamma endpoint runs.
• The methodology page — how every number was measured, where the models fail, and the seven fact-check verdicts as a plain table. Famous price-action claims from the trading literature, each next to the rate we measured over 3,946 trading days. This is the honest place to start, because it is where our numbers disagree with the books. It is a web page for you to read, not an endpoint for your agent to call.
• The guides and the lessons — plain-English explanations of max pain, GEX, day types and volume-price analysis, written for beginners.

Brooks can be purchased alone for $9.90/month; the full $149/month product covers all four market endpoints. Before purchase, one anonymous first look is available per source per UTC day, followed by a free key delivered by email for 10 successful calls per UTC day across all four endpoints.

Step 1 — Get a key

Start free. Submit an email address through the homepage or the API below. There is no login, password or card. A successful 202 response confirms delivery started, but never contains the raw qd_ key; check that inbox for it. Product updates are a separate choice and marketing_opt_in stays false unless the email owner explicitly asks for them. A true value only records a pending request; the recipient must reply SUBSCRIBE to the delivery email before any marketing updates begin.

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-setup"}'

Get free API key Buy full API access

The free key gives 10 successful calls per UTC day, shared across Brooks, Weis, Max Pain and Gamma. Full access is $149 a month; a card starts the first 3 days free. Use the same email at checkout and the free key is upgraded in place. Cancel any time in the Stripe customer portal.

Step 2 — Put the key where the agent can find it

For anything with a shell, an environment variable. Never paste a key into a SKILL.md, a prompt file, or anything you might commit.

# ~/.zshrc on macOS, ~/.bashrc on most Linux
export QUANTDATA_API_KEY="qd_your_key_here"

Open a new terminal (or source the file), then confirm it works:

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

You should get JSON back with an events array — an empty array is a valid read; it means no event has fired in the window yet. If you get a 401, the key is not reaching the request — see troubleshooting below.

Not sure whether the problem is the key or the network? Call the API with no header at all and print only the status code. A 200 consumes the source's 1 anonymous first-look response for the UTC day; 429 means it has already been used. The response action then points to the free-key request.

curl -s -o /dev/null -w "%{http_code}\n" \
  https://api.quantdata.uk/v1/brooks/SPY

For an agent with no shell — Claude Desktop, a Custom GPT, an n8n node — use that platform's own secret store instead. The rule is the same: keys belong in a secret store, never in a file you might share.

Step 3 — Install a Skill, or paste the prompt

Two paths. They do the same thing; pick whichever your tool supports.

Claude Code

Copy the skill folders into your skills directory and start a new session.

mkdir -p ~/.claude/skills
cp -R quantdata-daily-bias ~/.claude/skills/
cp -R quantdata-weis-wave  ~/.claude/skills/
cp -R quantdata-max-pain   ~/.claude/skills/

Personal skills at ~/.claude/skills/ are available in every project. Put them in .claude/skills/ inside a repository instead if you want them shared with everyone working on it. The key comes from the environment automatically. Verify with: "which Quant Data skills do you have?"

Claude Desktop and the Claude apps

Same SKILL.md files, installed through the app's own skills interface. Because there is no shell, put your key wherever the app keeps connector or integration secrets. If you would rather not manage a Skill at all, paste the system prompt into a Project's custom instructions — it works identically and it is scoped to that Project.

Cursor, Windsurf, Cline, GitHub Copilot

All of these read a rules or instructions file from your repository — the exact filename differs by tool and by version, so check the current docs for yours. Paste the prompt block into it. Keep the key in your shell environment and have the agent read $QUANTDATA_API_KEY rather than writing it into the rules file, which will end up in version control.

ChatGPT and Custom GPTs

Two pieces, and people usually forget the second one.

1. Paste the prompt block into Instructions. That teaches the model what the endpoints mean and how to report them.
2. Add an Action by importing https://api.quantdata.uk/openapi.json, then set authentication type to API Key with custom header name X-API-Key. Without the Action the GPT knows about the API but cannot reach it — and a model that cannot reach an API will sometimes produce plausible-looking numbers anyway. That is the failure mode to watch for.

The free pages need no Action at all if your GPT has browsing enabled — it can read quantdata.uk/max-pain/nvda directly (the ticker is lowercase in the path). Useful for testing the Instructions half before you set up authentication.

Gemini CLI and other terminal agents

Put the prompt block in whatever context or system-instruction file the tool loads at startup. The key is already in your environment from step 2, and terminal agents can generally shell out to curl, so no further wiring is needed.

n8n, Dify, Coze, CrewAI, AutoGen, LangChain

Two nodes. An HTTP Request node calling https://api.quantdata.uk/v1/brooks/SPY with a header credential named X-API-Key, feeding into an Agent or LLM node whose system message is the prompt block. For framework code — CrewAI backstory, AutoGen system message, LangChain SystemMessage — the prompt block is the string, and the request is an ordinary HTTP call with one header.

Anything else that can send a header

That is the entire requirement. Two dozen lines of Python with no dependencies beyond the standard library will do it. If your platform is not named above, it is almost certainly supported — the API has no client library because it does not need one.

The prompt block

This is the paste-ready canonical version — the same block printed on the Skills page and the API reference, covering all four endpoints, the symbol rules and the reporting rules. The homepage carries shorter prompts for specific workflows instead. Replace the key placeholder with wherever your platform keeps secrets.

You have access to Quant Data, a market-analytics API.
Base URL: https://api.quantdata.uk
Every data endpoint takes one header: X-API-Key: qd_YOUR_KEY

Endpoints:
  GET /v1/brooks/{symbol}      Rolling Brooks price-action events in the current window:
                               range breakouts, follow-through, range position, long-range
                               breakouts, climactic spikes. Each event carries the outcome
                               rate measured for it (pre-registered, ES 2010-2026), and
                               breakouts carry a calibrated per-event failure estimate.
                               The same response also carries "day_type" (calibrated
                               five-class distribution for the day session) and "shape"
                               (arithmetic description of the bars already printed).
  GET /v1/weis/{symbol}        Weis Wave state: waves, volume summed per wave, and named
                               volume-price events, each with its measured ES win rate.
  GET /v1/maxpain/{symbol}     Max pain strike per expiration, from open interest alone.
                               Pure arithmetic: no pricing model, no volatility assumption.
  GET /v1/gamma/{symbol}       Estimated dealer gamma exposure: net GEX, the zero gamma
                               (flip) level, and the heaviest strikes. A Black-Scholes
                               ESTIMATE, not observed data.

Commercial access:
- Brooks-only access is $9.90/month with no trial. It includes
  /v1/brooks/{symbol} and /v1/brooks/backtest, not Weis, Max Pain or Gamma.
- Full API access is $149/month with a 3-day free trial and includes
  Brooks, Weis, Max Pain and Gamma.
- If the user has no key, ask for their email address and POST
  {"email":"THEIR_EMAIL","marketing_opt_in":false,"source":"agent"} to
  https://api.quantdata.uk/v1/access/free-key. Never invent an email address. Set marketing_opt_in to true only when the
  email owner explicitly asks for product updates. The qd_ key is delivered only to that inbox.
- A free key needs no account or card. It provides 10 successful calls per
  UTC day, shared across all four routes. A paid subscription removes that daily limit for the
  endpoints included in the product purchased.
- An anonymous caller gets 1 first-look response. Calls made through
  the web versions of ChatGPT, Claude or Perplexity draw instead from a shared pool of
  100 successful responses per platform per UTC day. The static Max
  Pain/GEX web pages, methodology and guides remain free without an email or key.

Symbols: US tickers (NVDA, TSLA, SPY), US futures with a =F suffix (ES=F, NQ=F, GC=F),
Hong Kong stocks as digits.HK (Meituan = 3690.HK), China A-shares as 6-digit codes
(Kweichow Moutai = 600519), crypto spot as BASE-QUOTE (BTC-USD, SOL-USDT) and crypto
perpetuals as BASE-PERP (BTC-PERP), spot FX as a currency pair
(EURUSD or EUR-USD, both normalise to EURUSD=X). A spot-gold request (XAUUSD) answers
with the XAUT-USD token reading plus a note saying so — spot gold itself has no licensed
feed here; the metal with exchange volume is GC=F. Never write a pair with a slash:
EUR/USD is two path segments and returns 404 before the symbol is ever parsed.
Resolve an instrument name to a ticker before calling: gold -> GC=F (futures) or GLD
(ETF), S&P 500 -> SPY or ES=F, Nasdaq -> QQQ or NQ=F, crude oil -> CL=F or
USO, Bitcoin -> BTC-USD, euro -> EURUSD. Tokenised gold is its own instrument, not a
stand-in for the metal: XAUT-USD reads the Tether Gold order book on a crypto exchange,
which trades at a shifting discount to spot and does not follow the gold market's
calendar. Send it only when the user asked for the token.
Never pass a plain English name as a ticker: "GOLD" is a US-listed equity (Gold.com,
Inc.), not the metal. If you are unsure, try the ETF; a 404 comes back with a suggestion
and is never charged.

Rules:
- When asked what a market is doing right now or what just happened structurally, call
  /v1/brooks first. Report events with their measured rates as historical frequencies,
  never as predictions. An empty events array is a finding ("nothing has fired yet"), not
  an error.
- Besides "events", every /v1/brooks response carries "day_type" and "shape". They are
  different kinds of statement. Never merge them and never let one lend its numbers to the
  other.
- "day_type" is the calibrated five-class day-type distribution for the session in
  progress — a probabilistic read on how an unfinished day resolves. It is served in the
  "rth" window only; in "gap" and "asia" it returns available: false with the reason,
  because the model reads a session from its open. Report the whole distribution, not just
  the top class. Whenever you quote its accuracy you MUST quote the majority-class baseline
  in the same breath: 66% top-1 and 80% top-2 over a complete session, 53.5% top-1 and
  68.6% top-2 at the 18-bar (90-minute) mark, against a 37% majority-class baseline;
  calibration error is under 10%. Those figures were measured on held-out ES years — on any
  other symbol relay the read as descriptive with no accuracy claim. Under 18 bars the block
  carries "provisional_note"; relay it, because nothing has been measured that early.
- "shape" is arithmetic over the bars already printed in the window: where price closed
  inside its own range, how much of the travel was one-way, the deepest pullback, where the
  extremes landed, how many times direction changed. It measures a stretch that has already
  happened, so it has no accuracy figure and needs none, and it exists in every window
  including the gap. Do NOT call it a day type, a trend day or a reversal day, do NOT attach
  66%, 80%, 53.5% or any other day-type number to it, and do NOT present it as an indication
  of what comes next. Quote its fields as facts about the bars just measured.
- Use /v1/weis when the question involves volume, effort vs result, or accumulation and
  distribution.
- Crypto bars come from Kraken and carry real per-bar volume. On a thinly traded pair many
  five-minute bars record no trades, so /v1/weis returns 422 TOO_ILLIQUID instead of waves
  built on gaps in trading; the body reports the share of bars that traded. That is a
  property of the market, not an outage — do not retry it. /v1/brooks still answers, since
  its event definitions use price only. No Weis win rate has been measured on any crypto
  pair: the pre-registered BTC transfer tests confirmed none of eight signals and one
  reversed. Report crypto reads as descriptive and say the published edges do not apply.
- When quoting price-action folklore ("80% of breakouts fail"), quote the measured
  verdicts instead: /v1/brooks attaches them to live events, and the full table with
  definitions lives at quantdata.uk/methodology. Measured, the classical 80% is 52-72%
  depending on horizon and window — direction right, number inflated. Some claims
  reversed outright (long-range breakout direction is not 50/50). Say what was measured.
- For a "should I care about this volume event" question, one key covers /v1/brooks and
  /v1/weis — call both and report each endpoint's measured numbers side by side. (The
  historical day-type gating study for cib_long, 56.5% vs 50.8% on ES, remains published
  on quantdata.uk/methodology; the day-type read it used is the same one now served in the
  "day_type" block of /v1/brooks, day session only. Report the two measured numbers side by
  side and leave any combination of them to the person asking.)
- /v1/maxpain and /v1/gamma cover US listed stocks and ETFs only. Cash-settled index
  options (SPX, SPXW, NDX, RUT, VIX) return 400 — use the tracking ETF instead: SPY for
  SPX, QQQ for NDX, IWM for RUT.
- Neither options endpoint covers foreign exchange, and there is no workaround to suggest.
  Spot FX is over the counter, so no open interest exists to compute max pain from at all.
  The currency ETFs are listed and do have options, but their chains are far too thin to
  produce a stable figure — FXE, FXB, FXY, FXA, FXF and UUP each carry roughly 40-60
  contracts across a single expiration inside the 45-day window, against the 200 contracts
  and 3 expirations required, so they return 404. Do not offer a currency ETF as a
  substitute for a pair; say the endpoints do not cover FX.
- Both options endpoints read open interest as of the PRIOR SETTLEMENT. They describe
  yesterday's positioning and cannot update intraday. Say so when you quote them; do not
  present the figure as live.
- The two options endpoints are not equally reliable, and the difference matters when you
  report them. Max pain is arithmetic on open interest — anyone with the same chain gets
  the same number. Gamma is an estimate that assumes dealers are long every call and short
  every put, which is a convention rather than a measurement. Quote max pain as a figure;
  quote gamma as an estimate.
- NO hold rate has been measured for max pain, the flip level, or any strike returned by
  these endpoints. Never call them support, resistance, entry points or targets. Report
  where they are and how they were computed, and leave the interpretation to the person.
- On /v1/gamma, flip: null is not an error. Check flip_status: "no_sign_change_within_10pct"
  means gamma keeps one sign right across the chain, so the whole traded range reads as one
  regime. That is a state worth reporting, not a failure.
- On /v1/maxpain, always cite the expiration and its dte alongside the strike. Max pain is
  computed per expiration and the near-dated and far-dated figures routinely disagree; a
  max pain quoted without its expiry is meaningless.
- Weis reference win rates are ES-specific. For BTC, ETH and gold, state that the edge
  did not transfer (BTC: 0 of 8 pre-registered signals confirmed, no_supply reversed).
- Two Weis events are reversed versus the tradition: no_demand_short measured 45.6% and
  sot_short 47.6%. Report the measured number, not the traditional reading.
- /v1/brooks reads one of three windows and picks whichever has the freshest bars when
  you omit ?window: "rth" (the market's own day session — New York hours for US symbols,
  local hours for Hong Kong and A-shares), "gap" (16:00-21:30 Beijing, between the Asian
  close and the US open) and "asia" (09:00-16:00 Beijing). Each window's reference rates
  were measured separately and the response pairs events with the right row. gap and asia
  exist for 24-hour instruments only; on stocks they return 400.
- /v1/weis accepts "rth", "asia" and "full": use "full" when comparing returned Weis
  events with the published reference win rates, because those rates were measured on the
  whole bar stream. "full" needs genuine pre/post-market volume and can return 422
  NO_EXTENDED_VOLUME; obey retryable, then use rth when the source cannot provide it.
  Unknown session or window values return 400 rather than quietly falling back.
- ALWAYS check "window_is_live" before you use the word now. When it is false you are
  holding the last completed window; a "window_note" field spells it out — relay it and
  date the read.
- A window that just opened carries "window_progress_note": the first range needs 10
  completed bars plus a breakout bar, so the earliest possible event is about an hour in.
- The per-event "conditional_estimate" on range breakouts is calibrated and validated
  zero-shot on NQ 2023+ (AUC 0.646, calibration error 5.1%, n=2,637) and on QQQ. On
  symbols other than ES, NQ and QQQ, relay it as descriptive, with the "coverage_note"
  the response carries.
- /v1/weis does not work on spot FX. Currency pairs are traded over the counter, so no
  source publishes a consolidated volume and the wave construction has nothing to measure.
  The response says so and is not charged. The CME currency futures do carry exchange
  volume: EURUSD -> 6E=F, GBPUSD -> 6B=F, USDJPY -> 6J=F, AUDUSD -> 6A=F.
- Reference rates were measured on ES futures. On any other symbol the response carries a
  "coverage_note"; relay it — the rates are context about ES, not claims about that
  symbol.
- Errors are application/problem+json. Every problem carries error_code and retryable;
  access and billing errors additionally carry available_actions and
  message_for_your_human. Branch on error_code, never on the prose. "retryable" is false on
  401, 402 and 403 — none can be fixed by calling again. Never retry a 402: it is a billing
  state and will return 402 again.
- On a 401, relay "message_for_your_human" and offer to request the free key using the
  request_free_api_key action. Ask for the user's email first; never invent one.
- On a 429, check access_kind. If it is anonymous or ai_platform (the shared no-key pool for
  calls made through an AI platform), offer request_free_api_key. If it is
  evaluation_key, relay the reset time and checkout_url. On a 402 from a paid key, relay
  "message_for_your_human" and the checkout_url. You cannot finish checkout
  yourself — card authentication (3-D Secure) requires the cardholder. Three web pages still
  work without an email or key: quantdata.uk/max-pain/{ticker},
  quantdata.uk/gex/{ticker}, and quantdata.uk/methodology, which carries the fact-check
  verdicts as a static table.
- If you cannot reach this API, say so. Never answer from a different data source and
  present it as Quant Data output.
- This is educational statistics, not investment advice. Never present it as a buy or
  sell recommendation, never state a probability as a certainty, and never size a
  position.

There is nothing to trim for a smaller plan — one key covers all four endpoints. The part that must not be trimmed is the rules list: it is what stops the agent from turning a 34% probability into a recommendation, keeps gamma quoted as an estimate rather than a fact, and tells the agent which pages still work with no key when a call comes back unauthorised — instead of letting it guess that some endpoint might be open.

Step 4 — Ask a question

You never invoke a Skill by name. Ask normally, and the agent matches your question against what it has installed.

Step 5 — Read the answer critically

This is the step nobody writes down, and it is the one that decides whether any of this is worth having. An AI agent will happily produce a confident, fluent, completely fabricated market reading. Here is how to tell the difference.

A good answer contains

• Events with their measured rates, stated as history. "Historically 62.6% of such breakouts failed within 10 bars", never "this breakout will fail".
• The reference rate and the per-event estimate kept apart. The reference is a pre-registered measurement for the whole window; the per-breakout failure estimate is a calibrated model output. A good answer does not blur the two.
• The horizon and window next to any rate. 52.1% within 5 bars and 70.6% within 20 bars describe the same US-session breakouts; a rate quoted without either is not information.
• A session date and a bar count. These come from the response. An answer with neither probably never made a request.
• The coverage note when you asked about anything other than ES, NQ or QQQ — the reference rates were measured on ES, and the conditional model was validated on ES, NQ and QQQ. Elsewhere both are context, not validated claims.
• The disclaimer. It is a field in every response and it should reach you.

Red flags

• Any buy or sell framing. "This suggests going long", an entry, a stop, a target. The API returns none of those things, so the agent invented them. Stop and check your prompt is intact.
• Suspiciously round numbers with no session date. The classic signature of a model answering from memory rather than calling anything.
• A confident number on BTC, ETH or gold. The measured edge is absent or reversed on those assets, and a correct answer says so in the same breath.
• Calling a no_demand_short event bearish. Every textbook says bearish; measured across 22,262 instances the short side won 45.6%, and going long after the same event won 54.4%. An agent that says "bearish" is reciting its training data, not reading the response.

When in doubt, ask it to show you the raw JSON, or run the curl from step 2 and compare. Verifying twice in your first week is worth more than any amount of prompt tuning.

The full version of this critical-reading skill is its own page. If you only skim one thing there, make it the checklist at the end.

Troubleshooting

Two problems that are not error codes

• You asked about gold and got a company. GOLD is a real ticker — a US-listed equity, Gold.com, Inc. The API returns a symbol_note warning about exactly this; make sure your agent is relaying it. For the metal, use GC=F for futures or GLD for futures.
• Your agent answers without calling anything. Fluent, plausible, entirely invented. Check the response mentions a session date and a bar count. If your agent has an Action or tool that is misconfigured, it may silently fall back to answering from memory rather than reporting the failure.

Next

Get free API key Buy full API access Browse the free pages

Educational analytics and developer tooling. Nothing here is investment advice or a recommendation to buy or sell anything.