Rate limits

This page covers how requests are counted against a key: the windows, the headers on every response, the 429 responses, and how the live stream and MCP tool calls are metered. Read it before you build a poller or open streams.

Limits by tier

Tier	Requests per minute	Requests per day	Concurrent streams
`free`	30	2,000	1
`trader`	120	50,000	2
`pro`	600	500,000	10
`institutional`	3,000	10,000,000	20

The stream column is the per-minute limit divided by 60, at least 1 and at most 20, counted per API instance.

How counting works

Windows are calendar windows, not sliding: the minute window is the current clock minute and the day window is the current UTC day. Both reset on the boundary.
Every request that authenticates is counted once, in both windows, before anything else happens. A request that then fails with 400 or 404 still counts. A 401 does not.
The daily quota is checked before the per-minute limit.
A request rejected with 429 has already been counted. Keep sending and X-RateLimit-Remaining stays at 0 until the window resets.
The server caches some responses for a few seconds or minutes (see Data freshness). The cache sits behind authentication, so a cached response counts like any other.

Headers

Every authenticated response carries five headers, including 429 responses, error responses after authentication, and the live stream’s response.

Header	Meaning	Observed
`X-RateLimit-Limit`	Requests allowed in the current minute.	`600`
`X-RateLimit-Remaining`	Requests left this minute, never below 0.	`597`
`X-RateLimit-Reset`	When the minute window resets, in Unix seconds.	`1789649100`
`X-Daily-Quota-Limit`	Requests allowed today (UTC).	`500000`
`X-Daily-Quota-Remaining`	Requests left today, never below 0.	`499938`

import os, urllib.request

req = urllib.request.Request(
    "https://api.routeur.app/v1/status",
    headers={"X-API-Key": os.environ["ROUTEUR_API_KEY"]},
)
with urllib.request.urlopen(req, timeout=20) as r:
    for name in ("X-RateLimit-Limit", "X-RateLimit-Remaining", "X-RateLimit-Reset",
                 "X-Daily-Quota-Limit", "X-Daily-Quota-Remaining"):
        print(name, r.headers[name])

X-RateLimit-Limit 600
X-RateLimit-Remaining 597
X-RateLimit-Reset 1789649100
X-Daily-Quota-Limit 500000
X-Daily-Quota-Remaining 499938

Over the limit

Status	Code	When	`Retry-After`
429	`daily_quota_exceeded`	The day’s requests exceed the quota.	Seconds until midnight UTC, plus one.
429	`rate_limited`	The minute’s requests exceed the limit.	Seconds until the next clock minute, plus one.
429	`too_many_streams`	The key already holds its concurrent streams on this instance.	`60`
503	`stream_unavailable`	The instance holds 1,000 streams or is shutting down.	`5`

The body is the standard error envelope:

{"error":{"code":"rate_limited","message":"This key's per-minute request limit was exceeded."}}

{"error":{"code":"daily_quota_exceeded","message":"This key's daily request quota is used up."}}

Wait for Retry-After seconds and retry the same request. Do not retry faster: the rejected request was counted, so a tight retry loop keeps the window full.

What counts as a request

Traffic	Cost
Any `GET /v1/...` that authenticates, including `/v1/status` and `/v1/openapi.json`	1
A request answered `401`	0
`/healthz`, `/readyz`, unknown paths	0
Opening a stream	1
Each minute a stream stays open	1, against both windows
MCP protocol messages (`initialize`, `tools/list`, `prompts/list`, `resources/list`)	0
An MCP tool call	The REST requests it makes: 1 for an endpoint tool, up to several for a composed tool. Each tool’s description on the tool reference states its cost.
Reading MCP resources `routeur://categories`, `routeur://topics`, `routeur://topics/{slug}`	1
Reading MCP concept resources and `routeur://openapi.json`	0

Streams

Opening a stream counts one request and returns the rate-limit headers with the 200.
An open stream counts one more request per minute. If the daily quota runs out while the stream is open, the stream sends an error event with code daily_quota_exceeded and ends.
A key may hold at most the number of concurrent streams in the tier table, per API instance. One more returns 429 too_many_streams with Retry-After: 60.
Streams end themselves after 45 to 55 minutes with a reconnect event; reconnecting counts as opening a stream.

See Live stream.

MCP tool calls

Tool calls run through the same REST handler as your own requests, so the per-minute limit and daily quota are shared between REST and MCP. Over a limit, the tool call comes back as a tool error (isError: true) whose text is built as GET <path> answered <status> <code>: <message>, followed by Retry after <n> seconds. when the REST response carried Retry-After. An observed tool error for a missing market shows the shape:

GET /v1/markets/kalshi/NOPE-NOPE answered 404 market_not_found: The market is not in the latest graph.

See Agents and MCP.

Polling guidance

Graph data changes once an hour, Kalshi quotes every 15 minutes and Polymarket quotes every 30; the freshness table lists every cadence and each endpoint’s server-side cache. Poll at that pace and cache on your side. For anything faster, use the live stream: it costs one request a minute however much it delivers.

Get started

Concepts

Agents & MCP

Webhooks

Reference

Limits by tier

How counting works

Headers

Over the limit

What counts as a request

Streams

MCP tool calls

Polling guidance

Limits by tier

How counting works

Headers

Over the limit

What counts as a request

Streams

MCP tool calls

Polling guidance

Related