VotePredictor — MCP server & API

US election forecasts with a public accuracy record. votepredictor.com · docs

This repo holds the MCP manifest and the integration docs. The server is hosted — there is nothing here to install or run.

MCP server

Remote, Streamable HTTP. Add the URL to any MCP client:

{
  "votepredictor": {
    "url": "https://votepredictor.com/api/mcp"
  }
}

For stdio-only clients:

{
  "votepredictor": {
    "command": "npx",
    "args": ["-y", "mcp-remote", "https://votepredictor.com/api/mcp"]
  }
}

Registry: com.votepredictor/elections

Tools

Tools take what a person actually says, not ids you have to look up first.

Related MCP server: telekash-mcp-server

JSON API

Free, no key, CORS open. Index: https://votepredictor.com/api/v1

GET /api/v1/race/{race_id}            e.g. 2026_SEN_OH, 2026_HOUSE_CA-12
GET /api/v1/chamber/{senate|house}
GET /api/v1/member/{bioguide_id}      e.g. P000197
GET /api/v1/ratings?office=ALL

curl https://votepredictor.com/api/v1/race/2026_SEN_OH

Three conventions

These are what make the data safe to quote. If you build on it, please carry them through.

Nothing is certain. p_dem_win is clamped to 0.01–0.99. An election that hasn't happened is never 0% or 100%, however safe the seat.

Provenance is always stated. source is polls or fundamentals. Nobody surveys Wyoming, so its forecast comes from state lean, incumbency and the national environment instead. Polls take precedence wherever they exist — measured against the seats holding both, the fundamentals model agrees with the poll-driven one on 15 of 16 safe races but only 3 of 7 competitive ones.

Some races publish nothing. Where a candidate has left the ballot, the endpoint returns forecast: null with a reason rather than a number, because the stored probability describes a matchup that no longer exists. A race we hold nothing for returns 404 — a different answer, deliberately.

Accuracy

Every forecast is frozen when its race resolves, so the track record can't be edited after the fact. The ratings board scores every model, market, forecaster and pollster on resolved races since 1998, adjusted for how hard each race was — including VotePredictor's own model, ranked in place alongside FiveThirtyEight, Cook and Sabato's Crystal Ball.

Attribution

Free to use. Please attribute "VotePredictor" and link the specific page. Probabilities change daily until a race is decided.

Issues and requests welcome here.