Get a key. Send one header.

One ceiling, shared. MCP tool calls and REST calls made with the same key draw on the same 20,000-an-hour counter, and the counter is per key — so keys are how you keep a bulk job from starving an interactive client. The one lower limit is the per-minute cap, 200 a minute per key: it applies to /v1 requests only, never to MCP or archive downloads.

Why the archive is not counted. The ceiling counts HTTP requests, and requests are the wrong unit for bulk files: a Parquet reader pointed at a URL fetches the footer and then every column chunk as its own Range request — thousands for one daily partition. So archive files are authenticated and recorded on every request, and charged against the hour on none. Hyperliquid files have no URL of their own: the API answers with a signed link per file (one ordinary request), valid 7 days and served from Cloudflare R2.

Download whole files anyway. curl -O pulls a partition once and curl -C - -O resumes an interrupted one; remote-reading re-fetches the footer and column chunks on every query, at a transfer rate capped per connection.

A backstop, not a meter. The ceiling is enforced approximately — each server process keeps its own count and re-syncs it on a short interval — so the boundary is soft in both directions. Design to stay under it, not to spend it to the last request. When you cross it, REST and MCP answer 429 with Retry-After: 60. Cross the per-minute cap and /v1 answers 429 with Retry-After: 1 — wait a second and carry on, or pace a bulk job to under 200 requests a minute.

Hyperliquid pages by keyset, never by offset. A list response carries nextCursor; pass it back verbatim as cursor until it is null or a page comes back empty. On most list routes limit defaults to 100 and goes to 5,000; each route's reference states its own bounds.

A cursor is bound to the query that issued it — reuse it after changing start / end or on another route and you get a 400 telling you to restart without one. A malformed cursor is also a 400 (pass back the value verbatim), never a silent reset to page 1.

Polymarket /events returns {"events": […], "has_more": …, "next_cursor": …}: feed next_cursor back as cursor while has_more is true. One call covers at most 24 hours. /settlements/bulk pages the same way, in a settlements envelope. Most other Polymarket routes return a bare array; the exceptions are /wallets/{address} (one profile object) and /archive/datasets and /archive/url (one object each). /v1/onchain does not page at all.

import os, requests

H = {"X-API-KEY": os.environ["VIKE_API_KEY"]}
url = "https://data.vike.io/v1/hyperliquid/positions"

cursor = None
while True:
    params = {"coin": "BTC", "limit": 5000} | ({"cursor": cursor} if cursor else {})
    page = requests.get(url, params=params, headers=H, timeout=60).json()
    rows = page["positions"]          # the list key differs per route
    if not rows:
        break
    handle(rows)
    cursor = page.get("nextCursor")   # null on the last page
    if cursor is None:
        break

wss://data.vike.io/v1/stream pushes a bar frame each time a Hyperliquid candle closes, for every subscription you hold. Connecting, reconnecting and subscribing are not counted against the hourly limit, and neither are the frames you receive.

• Subscribe with {"action": "subscribe", "symbol": "BTC", "interval": "1m"}; optional market (perp | spot) and dex. unsubscribe takes the same fields.
• Intervals: 1m 5m 15m 30m 1h 4h 1d. Symbols are Hyperliquid assets (BTC, HYPE, @107).
• Frames: {"type": "bar", …} with ts (epoch ms), OHLCV, closed: true and a seq that rises by one per frame for each subscription — a jump means frames were dropped, but seq means nothing after a reconnect. Find a gap by ts: bars are one interval apart, so a bigger step is a range to check against REST (a bucket with no trades has no bar); a {"type": "ping"} heartbeat (answer {"action": "pong"}); and {"type": "error", …} with code bad_subscription, or rate_limited when the connection already holds 20 subscriptions, the hourly limit is reached or the stream is at capacity.
• A refused credential closes the socket with code 4401 before the handshake completes — most clients report it as an HTTP 403 on the upgrade.
• Backfill a gap from REST: https://data.vike.io/v1/hyperliquid/ohlcv serves the same bars.

Server: https://data.vike.io/mcp. Sign-in (OAuth) opens vike.io; an X-API-KEY header works too. Same data, same key, same hourly budget as REST.

• Claude Desktop and other OAuth clients need no key: add the URL, sign in once in the browser tab that opens, and approve. Connected apps are listed — and revocable — on your API keys page.
• Everything else sends X-API-KEY, exactly as on REST.

The spec is public at vike.io/data/openapi.json — no key needed, the same document the API publishes. The copy on the API host, https://data.vike.io/v1/openapi.json, and Swagger UI at https://data.vike.io/v1/docs do need a key: the credential check covers the whole /v1/ prefix on data.vike.io.

It documents every route's parameters. Response shapes are not modelled in it — they are on the hand-written reference at vike.io/docs/data, read from real responses.