How To Use Claude To Analyze Kalshi Markets (Without Copy-Pasting Numbers)

There is a specific kind of tedium every prediction-markets trader knows. You are reading about a market. You have a view. Now you tab out to a calculator, type in the price, type in your estimate, read a number, tab back, and lose the thread of what you were thinking about.

Multiply that by nine markets on a Sunday and the friction stops being trivial. It quietly changes which markets you bother to check.

So we removed the tab-out. Claude can now call our quant desk directly — 22 tools, sitting behind the same models that run our public tool pages — and answer in the same conversation where you are already thinking.

What this actually is

It's an MCP server. MCP is the open standard for letting an AI host call external tools, and ours is hosted — a URL, not a download. You point Claude at it once and the tools appear.

The important part is what the tools are. Claude is good at reading a situation and terrible at arithmetic it does in its head. Ask a language model "what's my edge at 62 cents if I think it's 71 percent" and you will get a confident number that is sometimes wrong. Ask a model that can call an EV function and you get the same answer every time, because a deterministic calculator computed it.

The model brings judgment. The tools bring the math. Neither should be doing the other's job.

Connect it in one command

Claude Code:

claude mcp add --transport http predictionmarketspicks https://predictionmarketspicks.com/api/mcp/mcp

Claude desktop or web app: add this URL as a custom connector:

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

That's the whole setup. No install, no local process, no key for the free tools. It's a standard remote MCP server, so the same URL works in ChatGPT and Cursor — we're on the official MCP registry as com.predictionmarketspicks/quant.

The first thing to ask

Start with the question you'd actually have. Here's a real exchange, with real output from the live server:

> You: A Kalshi contract is trading at 62 cents. I think it's closer to 71 percent. Is there an edge, and what should I size?

Claude calls calculate_ev and gets back:

{ "edge_pct": 14.52, "signal": "BUY",
  "interpretation": "Market is 14.5% underpriced vs. your estimate" }

Then it calls kelly_size against a $2,000 bankroll:

{ "stake_dollars": 236.84, "fraction": "half",
  "full_kelly_pct": 23.7, "applied_fraction_pct": 11.8,
  "rating": "optimal" }

A 14.5% edge, and half-Kelly says $236.84 — about 11.8% of the roll, against a full-Kelly number of 23.7% that almost nobody should actually take. You got the sizing discipline for free, in the same breath as the edge.

Note what didn't happen: Claude didn't estimate either number. It routed both to a function.

More prompts worth stealing

• "Convert +450 to an implied probability, then tell me what price that is in cents." — convert_probability, the one everyone gets backwards.
• "I thought this was 60%. Two new polls came in. Update me properly." — bayes_update does the posterior instead of vibes.
• "How often does a favorite at this price actually resolve yes?" — base_rate_gap compares your number to the historical base rate, which is where most bad positions die.
• "Price this three-leg same-game combo." — combo_edge handles the correlation that naive multiplication ignores.

What's free and what isn't

Fifteen of the 22 tools answer anonymously, no key, no account: every calculator, and the full NFL fantasy draft board (draft_board, best_available, player_outlook, compare_players, sleepers_and_busts, who_do_i_draft, adp_market_gaps).

Six need a Pro key, because they read live model output rather than computing from your inputs: cross-platform gap scanning, macro pulse, and our commodity and NFL edge models. One — edge alerts — is free on a 24-hour delay and real time with a key.

Everything the server returns is derived output: edges, tiers, model probabilities, sizing, and a link back to the tool page so you can check the working. It's analysis, not a data feed.

The honest limits

It won't make you right. These tools compute the consequences of your estimate. If your 71% is wrong, a very precise edge calculation on a wrong number is still a losing position. The base-rate tool exists specifically to argue with you about this.

Claude can still misread the setup. It's the model that decides which tool to call and what to pass it. It is good at this and not perfect. If a number looks strange, ask it to show you the tool call — the inputs are usually where the problem is.

Probability inputs are percentages. Pass 71, not 0.71. If an answer comes back wildly negative with a confident SELL, that's almost always what happened.

Why we built it this way

Prediction markets reward people who do a small amount of arithmetic consistently, and punish people who skip it because it was annoying. Most traders know the math. Very few run it every single time, on every market, because tabbing out has a cost and that cost compounds into skipped checks.

Putting the desk inside the conversation is our answer. Ask the question where you're already thinking about it, and let the calculator be the thing that's always right.

The endpoint is live now: https://predictionmarketspicks.com/api/mcp/mcp. Full tool list and per-client setup on the MCP page, and if you'd rather click than prompt, every one of these runs on its own tool page too.

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