Hyperliquid market data

All reads are POST /info with a JSON body; no key, no signing. Market data is usually read from mainnet even when the desk trades on testnet, because testnet prices and books are thin; say which network a figure came from. Every figure the desk reports carries source (request type), network and UTC time.

BASE=https://api.hyperliquid.xyz            # or https://api.hyperliquid-testnet.xyz
hl() { curl -sS -m 15 -X POST "$BASE/info" -H 'Content-Type: application/json' -d "$1"; }

Python header (SDK):

from hyperliquid.info import Info
from hyperliquid.utils import constants
info = Info(constants.MAINNET_API_URL, skip_ws=True)     # TESTNET_API_URL for testnet

Read-only snapshot contract

Run python3 scripts/opening_bell.py --schema or python3 scripts/desk_doctor.py --schema to inspect their machine-readable contracts without file or network reads. With --json, successful output carries schema_version: 1 and mode: read-only; invalid arguments return a JSON error with exit 2, while unavailable snapshot data or failed doctor checks exit 1. Decimal quantities remain strings. A depth floor is still a floor: inspect each bid_complete/ask_complete flag, and do not promote a doctor's WARN or offline SKIP to trading readiness.

python3 scripts/opening_bell.py --coin ETH --json
python3 scripts/opening_bell.py --coin kPEPE --json
python3 scripts/opening_bell.py --coin xyz:TSLA --json
python3 scripts/opening_bell.py --dex xyz --coin TSLA --testnet --json

A HIP-3 snapshot requests metaAndAssetCtxs with the explicit DEX and l2Book with the canonical prefixed coin from that universe. It retains the metadata's spelling and reports coin, dex, precision and leverage metadata; unknown, ambiguous or delisted markets fail before the book request. Metadata fields absent from the response are null, never invented. --testnet selects the official testnet origin; an explicit different --base-url is labelled custom and included in the source. Read-only discovery does not enable HIP-3 execution or transfer collateral; the desk's opt-in and ticket rules still apply.

Prices

hl '{"type":"allMids"}' | jq '{BTC, ETH, SOL}'                          # mid per coin, strings

allMids falls back to last trade when the book is empty. For mark, oracle and mid together use metaAndAssetCtxs below. Python: info.all_mids().

Market metadata, funding, open interest, volume

hl '{"type":"metaAndAssetCtxs"}' | jq -r '
  .[0].universe as $u | .[1] | to_entries[] | . as $e | $u[$e.key] as $m
  | select($m.name == "BTC" or $m.name == "ETH" or $m.name == "SOL")
  | [$m.name, $e.value.midPx, $e.value.markPx, $e.value.oraclePx, $e.value.funding, $e.value.openInterest, $e.value.dayNtlVlm, $e.value.premium, $m.maxLeverage, $m.szDecimals] | @tsv'

Fields per asset (same order as meta.universe): midPx, markPx, oraclePx, funding (hourly rate as a decimal: 0.0000125 = 0.00125%/h), openInterest (coin units), dayNtlVlm (24h USD volume), premium (impact bid/ask versus oracle, the input to funding), prevDayPx, impactPxs. Universe fields: name, szDecimals, maxLeverage, marginTableId, onlyIsolated/marginMode, isDelisted.

Python: meta, ctxs = info.meta_and_asset_ctxs().

Derived numbers the desk uses (show the formula): OI notional = openInterest x markPx; annualised funding = funding x 24 x 365; 24h change = markPx / prevDayPx - 1.

Margin tiers (max leverage by notional) come from meta:

hl '{"type":"meta"}' | jq --arg c BTC '(.universe[] | select(.name==$c)) as $u | ($u.marginTableId // $u.maxLeverage) as $id
  | {name:$u.name, maxLeverage:$u.maxLeverage, marginTableId:$id,
     tiers: (if $id < 50 then [{lowerBound:"0.0", maxLeverage:$id}]
             else ((.marginTables[] | select(.[0]==$id) | .[1].marginTiers) // [{lowerBound:"0.0", maxLeverage:$u.maxLeverage}]) end)}'

Ids below 50 are single-tier tables whose max leverage equals the id, and they are not listed under marginTables, so the snippet synthesises that tier; ids of 50 and above are looked up.

Order book and depth

hl '{"type":"l2Book","coin":"ETH"}' | jq '{time, bids: .levels[0][:5], asks: .levels[1][:5]}'

Up to 20 levels per side; each level is {px, sz, n} (n = number of orders). Optional nSigFigs (2-5) aggregates price levels; mantissa (1, 2 or 5) only with nSigFigs: 5.

Twenty levels is a page, not the book. On a liquid perp those levels stop a few bps from the mid - around 8 bps on ETH and under 3 bps on BTC at normal spreads - so a 25 bps band summed from the default response is whatever the page happened to contain, not the depth within 25 bps. Read the reach before quoting a band: when the side came back with 20 levels and the furthest one is nearer than the band, the number is a floor. To reach further, re-request with nSigFigs: 4, which buckets prices coarsely enough to push 20 levels out to roughly 20-25 bps on a major perp (measured live: 82 bps ETH, 25 BTC, 24 HYPE, 20 SOL), and drop to nSigFigs: 3 when even that stops short. Read each band off the finest page that reaches it - the coarse page moves the band edges by up to one bucket, and its top of book is not the real one - and say which page a figure came from. scripts/opening_bell.py is that ladder in code.

Depth within a band, the way the Risk Manager and Execution Trader want it:

hl '{"type":"l2Book","coin":"ETH"}' | jq '
  (.levels[0][0].px|tonumber) as $bb | (.levels[1][0].px|tonumber) as $ba | (($bb+$ba)/2) as $mid
  | def within(side; bps): [side[] | select((((.px|tonumber) - $mid) | fabs) / $mid * 10000 <= bps) | .sz|tonumber] | add // 0;
    def reach(side): (((side[-1].px|tonumber) - $mid) | fabs) / $mid * 10000;
  {mid: $mid, spread_bps: (($ba-$bb)/$mid*10000),
   levels: [(.levels[0]|length), (.levels[1]|length)], reach_bps: [reach(.levels[0]), reach(.levels[1])],
   bid_5bps: within(.levels[0]; 5), ask_5bps: within(.levels[1]; 5),
   bid_10bps: within(.levels[0]; 10), ask_10bps: within(.levels[1]; 10),
   bid_25bps: within(.levels[0]; 25), ask_25bps: within(.levels[1]; 25)}'

reach_bps is the answer's own scope: any band wider than it, on a side that returned 20 levels, is a floor and is quoted as >= size. Two bands reporting the same total is that cut-off, not a flat book.

Expected slippage for a size: walk the relevant side accumulating sz until the target size is reached; report the volume-weighted price versus mid in bps. If the size exceeds the visible 20 levels, say "beyond visible depth". Python: info.l2_snapshot("ETH").

Recent trades: hl '{"type":"recentTrades","coin":"ETH"}' (public prints: px, sz, side, time).

Candles

END=$(date +%s000); START=$((END - 48*3600*1000))
hl "{\"type\":\"candleSnapshot\",\"req\":{\"coin\":\"ETH\",\"interval\":\"1h\",\"startTime\":$START,\"endTime\":$END}}" \
 | jq -r '.[] | [.t, .o, .h, .l, .c, .v, .n] | @csv'

Fields: t open time ms, T close time ms, o h l c strings, v base volume, n trade count. Intervals: 1m 3m 5m 15m 30m 1h 2h 4h 8h 12h 1d 3d 1w 1M. Only the most recent 5000 candles per market and interval exist on the API, so history depth depends on the interval: about 3.5 days of 1m, 208 days of 1h, 2.3 years of 4h, 13 years of 1d. Choose the interval to fit the history you need; requests for older candles return nothing. Python: info.candles_snapshot(coin, interval, start_ms, end_ms).

For the strategy lab, use a research bundle instead of saving an unverified CSV:

cd /workspace/hypergrok
python3 scripts/research_dataset.py --schema
python3 scripts/research_dataset.py --coin ETH --interval 4h \
  --start 2026-09-01T00:00:00Z --end 2026-10-01T00:00:00Z \
  --with-funding --output /workspace/trading-desk/data/eth-september --json
python3 scripts/research_dataset.py --verify /workspace/trading-desk/data/eth-september --json

Choose explicit aligned UTC boundaries for the question, with end no later than collection time. The tool uses [start,end), closed candles only, at most 5,000 requested slots and fixed intervals from 1m through 1d (excluding 3d, 1w, 1M). With funding, boundaries must also be whole hours. It records canonical perp/DEX identity, origin/network, requests, decoded responses and retrieval times in sources.json; saves observed CSV rows; and writes a hashed manifest with complete, incomplete or unavailable coverage. It refuses an existing output directory. --schema and --verify do not use the network.

Return the bundle path, manifest hash, requested range, coverage gaps and offline verification result. Exit zero requires complete coverage, including funding if requested; an intact incomplete bundle still exits nonzero. No gap is filled, no absent rate becomes zero, and conflicting duplicates fail. Current metadata may include a delisted perp for historical research, flagged in the bundle; this does not enable trading it. A selected HIP-3 prefix/--dex, testnet or custom origin remains explicit. Spot is not supported by this collector.

See Research datasets for request bounds, pagination and handoff details. Replay and hashes establish internal consistency, not source authenticity, historical universe completeness, strategy edge or Risk PASS. Never describe a requested year as a collected year until the coverage proves it.

Funding history and predictions

START=$(( $(date +%s000) - 7*86400000 ))
hl "{\"type\":\"fundingHistory\",\"coin\":\"ETH\",\"startTime\":$START}" | jq -r '.[] | [.time, .fundingRate, .premium] | @tsv'
# venue, raw rate, its interval in hours, rate per hour, next funding
hl '{"type":"predictedFundings"}' | jq -r '.[] | select(.[0]=="ETH") | .[1][] | select(.[1]) | [.[0], .[1].fundingRate, (.[1].fundingIntervalHours // "?"), (if .[1].fundingIntervalHours then (.[1].fundingRate|tonumber) / .[1].fundingIntervalHours else "?" end), .[1].nextFundingTime] | @tsv'

fundingHistory returns hourly rates (up to 500 per call; paginate from the last returned timestamp, retaining overlap and checking duplicates). For research, use research_dataset.py --with-funding: actual millisecond timestamps are preserved and coverage is checked by UTC hour bucket. Requests are paced; --max-pages bounds funding collection (default 100, range 1-250). A page limit, missing hour or failed read is not a complete dataset.

predictedFundings compares venues, and the venue rates are over different periods, so never compare them raw. Every coin carries the same three slots in the same order - BinPerp, HlPerp, BybitPerp - but match on the venue name rather than the position, and each slot is either a payload or null when the coin is not listed on that venue. Divide fundingRate by the payload's own fundingIntervalHours to get a per-hour rate. HlPerp is always 1; the CEX venues are 4 or 8 depending on the coin, not a fixed 8: on mainnet on 2026-08-28, BinPerp was 4h for 126 coins and 8h for 63, BybitPerp 4h for 118 and 8h for 69. BTC, ETH and DOGE are all 8h, so an assumed 8 looks right on the majors and is wrong by 2x across most alts. A few BinPerp payloads (19 that day, including TON, MKR and IP) omit fundingIntervalHours entirely; treat the interval as unknown and say so rather than defaulting it.

Funding is paid every hour at size x oracle price x hourly rate; longs pay shorts when positive. Python: info.funding_history(coin, start_ms).

Rates plus candle closes do not reconstruct exact historical USD payments: that needs oracle prices and the position at payment time. Label a candle-price proxy as an assumption, never an observed funding cash flow.

Spot markets

hl '{"type":"spotMeta"}' | jq '.universe[] | select(.name=="PURR/USDC" or .name=="@107")'
hl '{"type":"spotMetaAndAssetCtxs"}' | jq '.[1][:3]'

Spot pairs are named PURR/USDC or @<index> on the API (the app shows HYPE/USDC); tokens: [base, quote] indexes into spotMeta.tokens. A pair's asset id for orders is 10000 + universe index; its size decimals are the base token's szDecimals. Ids differ between mainnet and testnet.

HIP-3 builder perps

Other perp dexs exist beside the main one: hl '{"type":"perpDexs"}' lists them; coins are named dex:COIN, and meta, metaAndAssetCtxs, clearinghouseState accept "dex": "<name>". The desk uses the default dex unless the user says otherwise.

Brief format

Use the block in agents/market-analyst.md: sources and time on the first line, then facts, derived, read, unknown, next.

Rate limits

REST shares an IP budget of 1200 weight per minute. allMids, l2Book, clearinghouseState, orderStatus, spotClearinghouseState and exchangeStatus cost 2; userRole costs 60; other documented /info requests cost 20. History/list endpoints add weight per 20 returned items, and candleSnapshot per 60 candles. Batch questions, pace pagination, and prefer hyperliquid-websocket for continuous data. HTTP 429 means back off; see the official rate-limit page before sizing a collector.

Pitfalls

• Reporting a number without the request type, network and UTC time.
• Treating funding as an 8h or daily rate; it is hourly.
• Comparing predictedFundings venues without dividing each by its own fundingIntervalHours, or assuming the CEX venues are 8h when most coins are 4h.
• Reading a book once and calling it "the depth" ten minutes later.
• Forgetting spot @index naming and base-token szDecimals.
• Expecting deep history at fine intervals; only the latest 5000 candles per interval exist, so pick the interval to fit the lookback.

HIP-3 equities while the underlying market is closed

• Price reference: oraclePx is the deployer's oracle reference, and prevDayPx is the API's previous-day reference. Neither field establishes the underlying share's official exchange close. Label markPx / prevDayPx - 1 as the perp's 24h change against prevDayPx, never "since Friday's close". An exchange-close comparison needs a separately sourced official close, its session/date and timestamp; otherwise report it unavailable. Name the price field, DEX-prefixed coin, network and UTC observation time.
• Mark/oracle and liquidity: Report (markPx / oraclePx - 1) x 10000 in bps only with a positive oracle and both fields present, alongside dayNtlVlm. This observed divergence is distinct from the API's impact-price premium used in funding. Zero or thin volume warrants checking the timestamped l2Book and recentTrades; a large divergence alone proves neither an executable price nor a stale mark. Do not infer an underlying stock move from it.
• Market identity: The same ticker on two DEXs represents separate markets with their own books, oracle definitions and funding settings. Query metaAndAssetCtxs with the explicit dex, and fundingHistory with the canonical dex:COIN; retain that prefix in every comparison. A ticker alone does not identify the funding rate.
• Funding window: For a weekend or other closed-market question, read several completed hourly fundingHistory rows with their UTC times, rather than presenting one current funding value as the period's rate. Show the hourly decimals and any simple annualisation (rate x 24 x 365, multiply by 100 for percent) separately; annualisation extrapolates that observation and predicts no future payment or return. Retain missing hours as gaps.
• External zero rates: Preserve a venue's reported 0 and distinguish it from null, missing or stale data. Check the venue's documented schedule and field semantics before labelling it actual zero accrual, a non-accruing interval or unreported funding. If that evidence is unavailable, say "reported zero; accrual/reporting status unverified". Do not replace every weekend zero with "no accrual" or compare rates whose intervals are unknown.