Draft night has a familiar posture: the draft room in one tab, a rankings site in another, an ADP page in a third, and a group chat calling you slow. By round 6 you've stopped cross-referencing anything and you're drafting on vibes and last year's box scores.
Here's the version we built instead. Your AI assistant — Claude or ChatGPT — calls our 2026 projection board directly, mid-draft, with the players already taken removed. You keep talking to the room. The model keeps the receipts.
All of it free to call, no account. Setup is one URL.
Connect it
Claude Code:
claude mcp add --transport http predictionmarketspicks https://predictionmarketspicks.com/api/mcp/mcpClaude app or ChatGPT: add that same URL as a custom connector (ChatGPT wants developer mode enabled, a paid-plan feature).
Seven draft tools come live immediately: draft_board, best_available, who_do_i_draft, player_outlook, compare_players, sleepers_and_busts, adp_market_gaps. Half-PPR by default; standard and full PPR by asking; platform presets for Yahoo, ESPN, Sleeper, NFL, CBS, Fantrax, and best-ball defaults for DraftKings and Underdog.
On the clock: a real round-4 spot
Real exchange, real output from the live server. Pick 40, 39 players gone, roster so far: Bijan Robinson, Tee Higgins, Trey McBride.
> You: Pick 40's on the clock, here's who's gone. I have Bijan, Higgins, McBride. Who's the play?
The assistant calls best_available with the drafted list and gets back, in part:
{ "best_value": { "player": "Alec Pierce", "team": "IND", "position": "WR",
"adp": 62, "edge": 12, "value": "SLEEPER" },
"available": [
{ "player": "Zay Flowers", "team": "BAL", "position": "WR",
"proj_total": 271.2, "pos_rank": 5, "adp": 30, "edge": 10 },
{ "player": "George Pickens", "team": "DAL", "position": "WR",
"proj_total": 214.8, "pos_rank": 14, "adp": 20, "edge": -5 } ] }Two different answers to two different questions, which is the point. Zay Flowers is the pick: the model projects him WR5 (271 half-PPR points) and he's still sitting there in round 4 at a round-3 ADP. Alec Pierce is the note for later: the single biggest value left on the whole board (+12 spots of model edge over his cost), but at ADP 62 he'll still be there in two rounds — drafting him at 40 pays retail for a discount player.
That distinction — best pick now versus best value left — is exactly what you stop tracking by round 6 on your own.
ADP arbitrage: same player, different price
Prediction-market habit, applied to draft capital: the same asset trades at different prices on different venues. adp_market_gaps surfaces the spread. Real output, top of the current board:
{ "players": [
{ "player": "Travis Hunter", "position": "WR",
"markets": { "consensus": 185, "espn": 101 }, "spread": 84 },
{ "player": "T.J. Hockenson", "position": "TE",
"markets": { "consensus": 187, "espn": 114 }, "spread": 73 },
{ "player": "Jaxson Dart", "position": "QB",
"markets": { "consensus": 126, "espn": 62 }, "spread": 64 } ] }Travis Hunter costs a 9th-round pick on ESPN and a 16th-round pick on consensus — an 84-slot spread on the identical player. If you're in an ESPN league, that's the market telling you where not to reach; everywhere else, he's nearly free. Hockenson and Dart, same story at TE and QB.
Sleepers, busts, and the round-8 minefield
sleepers_and_busts filters by round. Round 8, current model, verbatim flags:
- Xavier Worthy (KC) — ADP 96, model has him WR60. Edge -15, tagged
CONFIRMED-BUST. - Chris Godwin Jr. (TB) — ADP 85, model WR59. Edge -19, tagged
BUST. - Jordan Addison (MIN) — ADP 91, model WR30, projected 175 points. Edge +12, flagged as the value in the round.
One round, two landmines and a discount, and the model will happily show its work on any of them — player_outlook gives the floor/ceiling range, compare_players puts two names side by side when you're torn.
And if you just want the big board: draft_board returns it ranked, where the model currently makes Jaxon Smith-Njigba the overall WR1 by projection (393 half-PPR points) at an ADP of 7 — the kind of disagreement with consensus you want to know about before pick 7, not after.
The honest limits
It's a model, not a prophecy. Projections are our best blend of the numbers; they will be wrong about specific humans, because humans. Use the edge flags as an argument, not a verdict.
Tell it who's gone. best_available is only as current as the drafted list you give it. Feed picks in as they happen — one message every round or two is plenty.
Speed matters on a 60-second clock. The tools answer fast, but front-load: ask for sleeper targets and round plans before the draft, and save mid-draft calls for the picks that are actually close.
The anonymous lists are shallower than the full board. Every tool answers without a key, and player_outlook and compare_players are never capped — but ask for the board cold and you get the top 12, not all 50. Same for best-available (3) and sleepers/busts (3 of each). That's enough to draft off, and it's the honest default rather than a teaser. A free key lifts those to 50, 10 and 10 for an email — no card, no trial, no expiry.
Why we built this
We run a prediction-markets quant desk — the first one your AI agent can call — and a draft is just another market: prices (ADP), a model (projections), and edges where they disagree. The same discipline that grades a Kalshi contract grades a round-4 pick.
The endpoint is live: https://predictionmarketspicks.com/api/mcp/mcp. The full board and every draft tool also run on the draft page, setup details on the MCP page, and our NFL market work lives at /nfl.
Draft responsibly. Projections are arithmetic, not destiny — but arithmetic beats vibes by round 6.