Skip to main content
Routeur runs a remote MCP server at https://api.routeur.app/mcp. Any client that speaks the Model Context Protocol — Claude, Cursor, ChatGPT, the OpenAI Agents SDK, your own agent — can read every market on Kalshi and Polymarket, the relations proven between their contracts, the mispricings those relations expose, and strategies measured against settled history, without writing a line of HTTP.

Claude Code & Claude Desktop

One command, or one block of JSON.

Cursor

Add it to mcp.json and use it from Composer.

ChatGPT & the OpenAI Agents SDK

Hosted MCP tools, or a few lines of Python.

Any MCP client

Streamable HTTP directly, or mcp-remote for stdio-only clients.

​
What the server is

A request without a key is refused with 401 and a WWW-Authenticate: Bearer challenge, exactly like the REST API.

​
What it exposes

Every documented GET endpoint is a tool, generated from the same OpenAPI description this reference is built from. The tool is named after the operation (list_opportunities, get_market, list_relations), its arguments are that endpoint’s parameters, and its result is that endpoint’s JSON body, unchanged. When an endpoint is added to the API, the tool appears with it.On top of those, five composed tools answer a whole question in one call and return compact rows a model can actually read: find_opportunities, compare_venues, explain_relation, market_snapshot and search_markets. See the tool reference.
routeur://concepts/glossary, routeur://concepts/relations (how a relation is proven), routeur://concepts/leads-and-fees (what depth, executable and estimated fees mean), routeur://concepts/prices (asks, spread, volume, freshness), routeur://openapi.json, plus the live routeur://categories, routeur://topics and routeur://topics/{slug}.
scan_for_mispricings (sweep a category for prices that break a proven relation), brief_me_on_event (a sourced briefing on one event), backtest_price_band (what buying a price band returned in settled history). Each names the tools to use and the traps to avoid.

​
Keys, limits and metering

Your MCP calls and your REST calls are the same traffic on the same key.
  • A tool call counts as the API requests it makes. get_market costs one; market_snapshot, which reads five endpoints, costs five. Each tool’s description says what it costs.
  • Protocol messages — initialize, tools/list, prompts/list — are authenticated but not counted.
  • The per-minute limit and daily quota are your tier’s. Over either, the tool call comes back as a tool error naming rate_limited or daily_quota_exceeded, with the seconds to wait.
  • Reading routeur://categories, routeur://topics or a topic counts as one request; the concept pages and the OpenAPI resource cost nothing.
Keep the key server-side. In a desktop client it lives in that client’s config file; prefer an environment variable over pasting the key into a file you might commit.

​
Everything is read-only

Every tool is marked readOnlyHint and idempotentHint. Nothing on this server places an order, follows a market, posts anything or changes an account — Routeur is not a broker, exchange or adviser, and there is no write path today. User-scoped actions, if they arrive, will be a separate, clearly-marked surface behind OAuth, never a silent addition to these tools.

​
What an agent should know before it quotes a number

These are the mistakes models make with this data. The server states them in its instructions and its concept resources, but say them in your own system prompt too.
  • Prices are asks per contract that pays $1, as exact decimal strings. The spread (yes_ask + no_ask - 1) is what crossing costs; a wide spread is a thin book, not an edge.
  • A relation is proven, never guessed. No relation coming back means none was proven, not that two markets are unrelated.
  • A gap between venues is only money when both books have size and fees leave something. Only leads with depth.executable could be filled at a profit when the books were read.
  • Strategy and calibration figures are settled history, before fees, and are not a forecast.
  • Kalshi is read every 15 minutes, Polymarket every 30, and the graph is rebuilt hourly; every graph-derived answer names its graph_run.

​
Not using MCP?

https://routeur.app/llms.txt and https://routeur.app/llms-full.txt describe Routeur for agents in plain text, and the REST API is the same data over ordinary HTTP.