Connect ChatGPT To Live Prediction-Market Edge Data (One URL, No Install)

Most people use ChatGPT for prediction-market research the slow way: paste a price in, describe a situation, get an answer that sounds quantitative, and hope the model did the arithmetic right in its head. Sometimes it did. You will not reliably know which times.

There is a better wiring. ChatGPT supports remote MCP connectors — the same open standard Claude uses — which means it can call our quant desk directly. Twenty-seven tools, the same models behind our public tool pages, answering inside the conversation.

This is the setup guide, with real output from the live server.

One URL

In ChatGPT, custom connectors live behind developer mode (a paid-plan feature — Free accounts can't add them yet):

1. Settings → Connectors — enable developer mode under advanced settings.
2. Add a custom connector with this URL:

https://predictionmarketspicks.com/api/mcp/mcp

That's the entire install. It's a hosted Streamable HTTP endpoint — no download, no local process, no key to start calling. We're on the official MCP registry as com.predictionmarketspicks/quant, and the identical URL works in Claude and Cursor if you run more than one assistant.

The showcase: a combo that grades itself

The tool built for ChatGPT specifically is combo_edge. It ships as an MCP App — the result can render as a card inside the chat, not a wall of numbers.

Here's a real exchange, real output:

> You: Three legs at 62, 55 and 70 cents. Correlation-aware, I make the joint probability 28 percent. The platform is quoting me +320. Worth it?

ChatGPT calls combo_edge and gets back:

{ "verdict": "SMASH", "ev_pct": 17.6,
  "fair_american": "+257",
  "fair_band": { "low_american": 228, "high_american": 291 },
  "offered_american": "+320",
  "negative_correlation_trap": false }

Fair value on your 28% is +257, the acceptable band tops out at +291, and the venue is paying +320 — a 17.6% expected-value overlay, with the correlation-trap flag explicitly clean. That's a seven-tier verdict (SMASH / PLAY / LEAN / RISK / NO_VALUE / PASS / RUN) computed against the quote your platform actually pays, never against the naive product of the leg prices — which no venue pays, and which is the mistake that makes most multi-leg positions look better than they are.

Note the division of labor: you brought the correlation judgment (the 28%). The server brought the pricing math. ChatGPT just carried the conversation.

What "live" means here

The calculators are stateless — they compute from your inputs, free, instantly, no key. The live layer is different, and we'd rather explain it straight:

• edge_alerts is the feed our models write on Kalshi — weather, bitcoin, silver, gold, oil, mispricings. Every alert carries feed, tier (SPECULATIVE / MODERATE / STRONG), side, price in cents, model probability, edge in percentage points, and a market link. No key: the identical feed, delayed 24 hours, thesis stripped. That delay is deliberate — it lets you audit whether the signals were any good before paying for them to be fresh.
• The fantasy draft tools sit in between. Every one answers with no key, and player_outlook / compare_players are never capped. The list-shaped ones return a genuinely useful slice anonymously — the top 12 of the board, the top 3 available, 3 sleepers and 3 busts — and a free key lifts those to 50, 10 and 10 for an email, no card.
• A Pro key turns the feed real-time and unlocks the live model tools: find_arbitrage (cross-platform gaps), market_pulse (macro composite), commodity_edge, and the NFL edge models.

Everything returned is derived output — edges, tiers, model probabilities, sizing. Analysis, not a data feed, and every result links back to its tool page so you can check the working.

Prompts worth stealing

• "Any edge alerts on Kalshi in the last day? Weather only, MODERATE or better." — edge_alerts with filters.
• "Contract at 62 cents, I say 71 percent. Edge and half-Kelly size on $2,000." — calculate_ev then kelly_size, chained in one breath.
• "How often does a market priced like this actually resolve yes?" — base_rate_gap, the tool that argues with you.
• "Grade this quote against fair value before I touch it." — combo_edge, every time, no exceptions.

The honest limits

Probability inputs are percentages. Pass 71, not 0.71. The server rejects fraction-looking inputs with a correction rather than guessing at what you meant — a wrong-but-confident answer is worse than an error.

The delayed feed is genuinely delayed. If you ask for alerts on a quiet stretch, you may get zero back. That's the honest shape of a real feed, not a malfunction — and it beats a feed that invents signals to look busy.

ChatGPT still picks the tools. It decides what to call and what to pass. It's good at this and not perfect; if a number looks strange, ask it to show the tool call. The inputs are usually where the problem is.

Why this exists

We think of it as the first prediction-markets quant desk your AI agent can call. The people who win in these markets are the ones who run small arithmetic consistently — every market, every time. Wiring the desk into the assistant you already talk to removes the last excuse not to.

The endpoint is live: https://predictionmarketspicks.com/api/mcp/mcp. Full tool list and per-client setup on the MCP page, and the free feed is browsable at /alerts if you'd rather look before you connect.

Trade responsibly. Nothing here is financial advice — it's arithmetic, run consistently.