Connect your AI to Parlay Builder

This page is loading — it needs JavaScript. The written page below is served without it.

Parlay Builder publishes an MCP server — a small, read-only endpoint that an AI client can query directly. Add it once to Claude or ChatGPT and the assistant can pull tonight’s ranked prop board, how often a player has actually cleared a line, and the graded record of what the model has published, without you leaving the chat to look any of it up.

What an MCP connector actually is

MCP is the Model Context Protocol, an open standard Anthropic published in 2024 for connecting AI assistants to outside data. A “connector” or “MCP server” is one such source. The client asks it questions in a fixed shape and it answers with structured data instead of a web page, which is what lets an assistant reason over the numbers rather than scrape them out of prose.

What your assistant can ask for

“What’s on the NBA board tonight?”

pb_top_props — The ranked prop board for a league and date, filtered the way you would filter it on the site — league, market, minimum sample, how far out the game is.

“How often has Judge cleared 0.5 hits?”

pb_player_hit_rate — One player, one market, one threshold, measured across five windows — last 3, last 5, last 10, this season and last season — each with the games behind it, plus trend, consistency and current streak.

“Which of these three has cleared it most?”

pb_compare_props — Two to six props measured side by side on one window, so the comparison is like for like rather than three separate answers a reader has to line up.

“Who is steadiest on the MLB board?”

pb_most_consistent — Board rows ordered by the measured consistency index — how little a player’s output has moved around its own average — on the same thresholds the chat uses.

“Does any of this actually work?”

pb_track_record — The forward record: every prop the model has published since inception, graded against the final box score, broken out by league, market and confidence band.

“How much of tonight is covered?”

pb_board_coverage — How many of the slate’s games and props the board actually has measured history for, which is the number that tells you whether an empty answer means a quiet night or a thin feed.

“What does this number mean?”

pb_explain_metric — A plain definition of any figure the other six return — what it counts, over what window, and what it does not claim.

What it deliberately does not send

No sportsbook prices, and no book names
The data provider’s terms permit displaying their prices in a product and prohibit re-serving them as a source of raw data for others. Whether an authenticated MCP endpoint is the first thing or the second is genuinely unclear, so the endpoint does not test it: every posted price, bookmaker name and price age is stripped inside the response builder. What travels instead is whether a book has the selection up at all, and a link to the page on this site where the number is shown.
No grade, no edge figure, no projected return
Nothing in any response carries a rating or a number a model could read as a return. On the site those figures sit under a legend that says what they are and what they are not; handed to somebody else’s model there is no legend and no last hop left to correct it in. What travels is how often something has happened, over how many observations, and the board’s own ordering as a bare rank with no magnitude attached.
Nothing it can change, and nothing it can spend
Every tool is a read. The connector cannot place a bet, talk to a sportsbook, move money, see a payment method, or alter anything in your account. The key is scoped to one account and reaches no other subscriber’s data.
The threshold stays, and that is a deliberate exception
The line a rate was measured at does travel, because a hit rate without it is a fragment — “cleared this number in 15 of 22 games” stops meaning anything if the number is removed. A threshold on its own is not what an odds feed sells. That call is argued in writing rather than assumed, and it is the one judgment in the whole ruling.

Which AI clients can connect

Claude — supported
The client this was built and verified against. Settings → Connectors → “Add custom connector”, paste one URL. Available on Free, Pro, Max, Team and Enterprise; a free account is limited to one custom connector, and on Team or Enterprise an Owner adds it once for the organisation.
ChatGPT — supported
Works, behind Developer mode. Business, Enterprise and Edu get full tool access; Pro is limited to search and fetch actions, and a Free account cannot add custom connectors at all. On a managed workspace an admin may have to allow custom connectors first.
Gemini — not supported today
Not in the consumer Gemini app, which has no dialog for adding a custom MCP server — so there is no URL to paste and no instruction we could give that would work. The endpoint speaks Streamable HTTP, which is the transport Gemini Enterprise requires, so an administrator on that tier can add it as a custom MCP server data store.

Who it is for

  • People who already live in an AI assistant and would rather ask it than open a tab, read a board and retype what they found.
  • Anyone building their own research workflow — a local MCP client, an editor, a script — who wants measured history in a fixed shape instead of scraped out of a page.
  • Readers who want the sample size attached to every rate by default, because an assistant that answers from structured fields cannot quietly drop it the way a summary of a web page can.

The connector is included with the Pro plan and with the free trial. It is the one thing Core does not carry, and the plan check runs on the server every time a key is used rather than only when one is issued.

Setting it up

  1. Generate a connector key

    On the setup page, once signed in. One key per account, shown once at the moment it is created — we store only a hash of it and cannot show it to you again. Lose it and you revoke and regenerate, which is a thirty-second job.

  2. Paste one URL into your AI client

    The endpoint with your key already appended. That is the whole of the setup and the whole of the authentication: no OAuth fields to fill in, no config file to edit unless you want one. Treat the pasted URL like a password, because it carries one.

Questions about the AI connector

What is an MCP server, in one sentence?

A small endpoint that answers an AI client’s questions with structured data instead of a web page, using the open Model Context Protocol — so the assistant reasons over real fields rather than over whatever it managed to scrape.

Can my AI assistant see live sportsbook odds through this?

No, and that is deliberate rather than a gap. Every posted price, book name and price age is stripped from the response before it leaves, because our data provider permits displaying prices in a product and prohibits re-serving them as raw data. The connector tells your assistant whether a book has the selection priced and links to the page here where the number is shown.

Will it tell me what to bet?

No. No response carries a grade, a rating or a projected return — those fields are removed inside the response builder, so a leak is an error rather than a quiet regression. What comes back is how often something has happened, over how many games, and where the board ranked it. Deciding what that is worth is still yours.

Which AI clients can add it?

Claude, which it was built and verified against, and ChatGPT behind Developer mode. The consumer Gemini app has no way to add a custom MCP server, so it cannot; Gemini Enterprise can, added by an administrator as a custom MCP server data store.

Is it safe to give an AI assistant a key to my account?

Every tool is a read. The connector cannot place a bet, move money, see a payment method or change anything in your account, and the key is scoped to you — it reaches no other subscriber’s data. The key does travel in the URL you paste, because no hosted AI client can send a custom header, so treat that URL like a password and revoke it if it ever ends up somewhere it should not.

Do I need to be on a paid plan?

Yes. The connector is included with the Pro plan and with the free trial, and the plan check runs on the server on every call rather than only when a key is issued. The rest of the site — the board, the hit rates, the research pages — is still browsable without an account.

What happens if I lose my key?

You revoke it and generate another on the setup page. We store only a hash of a key and genuinely cannot show it to you a second time, so there is no “reveal” that would recover it. Revoking invalidates whatever client was using it, which is the point.

Odds and lines come from licensed data feeds and refresh on a schedule; hit rates, edges and grades are Parlay Builder’s own calculations.