What an AI agent gets
Four blocks, one path: connect, see, receive, know. Every number on this page is read at build time from the code that defines it — tool counts, tiers, start dates. None of it is typed in by hand, none of it ages unnoticed.
1 · Access
Two key kinds: OAuth connector (claude.ai, Claude Desktop, ChatGPT) and API key (REST, stdio, your own agents). Three tiers with call limits per hour and per day; prices live on the pricing page. arena_status tells the agent its tier, remaining quota and reset time at any moment — at no extra cost.
2 · Tool map
49 tools are in the default listing, 51 more in the extended set — 100 in all. The groups are Composites, Market, Assets, On-chain, Derivatives, Research; composites first, because one call there answers several single questions. Listed tools are the ones a client offers to the model. Every client reaches the rest through arena_call_extended: one call lists them, the same call runs them.
3 · What every answer carries
- Time
as_of · price_semantics— Every answer carries its measurement time and says whether the price is live, a daily close or a snapshot. - Provenance
provenance.grade · sources_used / expected— Composites state how many of their inputs were present today. Below half, there is no verdict. - Sample
n · window · benchmark— No verdict without a denominator: trade count, time window and benchmark window sit next to every figure. - Version
engine_version · reproducibility hash · supersedes— Knowledge Objects are immutable. A correction is a new object that points at the old one. - Robustness
dsr · is_distinguishable— Strategy verdicts are corrected for multiple testing; two numbers count as different only beyond the measured noise floor. - Licence boundary
disclosure · attribution— Licensed levels stay with the vendor. What is delivered is the shape: slope, rank, premium — with the source named in every answer.
Implied-volatility levels (DVOL, ATM IV per tenor) are not delivered through the API: they are licensed exchange data. What this endpoint delivers is their SHAPE and RANK — term-structure slope, contango/backwardation, percentile rank within our own history, and the volatility risk premium in percentage points. That answers "are options expensive, and did the expectation hold?" without passing on the price level itself. Realized volatility is available from arena_get_volatility_history; it is omitted here because premium + realized would reconstruct the licensed level.
4 · Limits the agent must know
- Point-in-time universes 2026-07-14 — From this day on, membership is recorded daily, delisted coins included. Before it, a universe backtest carries survivorship — and says so.
- Cycle score walk-forward 2026-04-17 — Since this day the Bitcoin cycle score is stored with that day’s knowledge. Earlier values are reconstructed and labelled as such.
- ETF flows: availability D+1 23:30 UTC — The source carries no publication timestamp. Measured from our own first ingest (daily 23:30 UTC): a US trading day D first appears on D+1 at 23:30 UTC — not at the US close of D. A point-in-time study should treat day D as known from D+1 at the earliest. Rows loaded by the initial backfill are excluded from the measurement.
- On-chain histories as reported — Values are as-reported: on-chain metrics can be revised retroactively, so this is not a point-in-time vintage.
Describes the product, does not grade it. No market figures, no recommendation — evidence an agent can check.