Skip to main content
The API and MCP server are in production today, grounding the chat analyst’s market answers. The discovery surface is already public: run curl https://noxint.ai/api/v1 right now, with no key, and the contract answers. Key minting and paid access open in beta after launch; credit packs are rolling out with launch. Launch status tracks exactly what is on.
Single calls prove an endpoint works. Workflows prove an agent can do a job. Everything below was run against the production gateway on 2026-07-16; responses are trimmed, never altered. Prices are a snapshot of that morning.

1. The positioning sweep

The question every derivatives desk asks at the open: where is positioning stretched right now? Two calls. Step 1: screen the core tier for extremes. rank with the extreme_count metric scores each asset by how many of its metrics sit in their own historical tails.
X-Credits-Charged: 5. The screen ranks only symbols with fresh inputs at scan time and reports its scanned and excluded counts in a coverage block, so the agent knows how wide the net was. Bitcoin Cash leads with three simultaneous extremes, one of them funding. The agent stores the canonical asset_id and drills exactly one domain. Step 2: drill the extreme.
X-Credits-Charged: 1. The read is unambiguous: shorts are paying 23% APR to stay short, the rate sits in the 9th percentile of its 90-day history and the 7th of the last year, all four venues agree, and the sign flipped only eight bars ago. Crowded, fresh, unanimous. Whatever the agent decides to do with that, it decided from ranked facts, and it knows to check back after the 16:00 funding settle. The whole sweep: 6 credits, six cents, two calls.

2. From ticker to filed facts

An equity dossier, starting from nothing but the word “apple”. Step 1: resolve, free. resolve?q=apple returns best: null, because “apple” is genuinely contested: Apple Inc. on Nasdaq, an OTC filer named Apple iSports, and two Solana tokens with the ticker APPLE. The ranked matches carry warnings instead of a silent guess. The agent takes the equity: eq:0000320193:aapl. Step 2: the situation. read?id=eq:0000320193:aapl returns the headline (price 327.50,marketcap327.50, market cap 4.84T, P/E "43.9" as an exact string), stamped filings_indexed_through: 2026-07-11. One credit. Step 3: the filed statements.
Revenue to the dollar, two quarters side by side, each tied to the accession number of the 10-Q it came from. One credit. Step 4: who is selling. events?types=insider returns the recent Form 4 tape plus a summary of the trailing 90-day window: zero open-market buys, one sale, net -116 shares across two unique insiders. One credit. The whole dossier: 3 credits. Ticker to SEC-grade facts for three cents, and every number in it carries its own provenance.

3. The monitoring loop

Live data agents die by polling: too fast burns money, too slow misses the move. The contract schedules the loop for you.
  1. read?id=crypto:bitcoin with an Idempotency-Key. The response carries next_update_at: 2026-07-16T09:45:19Z and an ETag.
  2. Sleep until next_update_at. Not a guess: the engine told you when the next change can exist.
  3. Re-request with If-None-Match. Unchanged data answers 304 and costs zero credits. Changed data bills one credit, and only then does your agent re-reason.
  4. If the request times out mid-flight, resend it with the same Idempotency-Key: transient failures release the hold, so the retry runs once and settles once.
A watch loop over one asset costs one credit per actual change in the data, not one per poll. Over MCP there is no conditional re-poll, so apply step 2 twice as strictly: cache the last result and never re-call before next_update_at.
Free to try today, before keys open: the docs and schema operations answer keyless, and resolve, coverage, and usage cost zero credits once you hold a key. The two jobs above total nine cents end to end.