Fantasy Football Draft Analyst (MCP Server)

A live draft assistant. Connect it to Claude, sync your ESPN or Sleeper draft board, and ask "who should I pick?" at every turn. It answers with a recommendation, the reasoning, and the odds each player survives to your next pick.

Tested end to end against live data: 631 players (551 veterans, 80 rookies) modelled from 247,284 plays across five seasons, priced against real 2026 preseason consensus.

Quick start

git clone https://github.com/OWNER/ff-draft-mcp.git
cd ff-draft-mcp
pip install -e .

# One-time data build (~3-6 min; downloads 5 seasons, then caches)
python setup_data.py

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "fantasy-draft": {
      "command": "python",
      "args": ["-m", "ffdraft.server"],
      "env": {
        "FFDRAFT_SEASON": "2026",
        "ESPN_SWID": "optional-for-private-espn-leagues",
        "ESPN_S2": "optional-for-private-espn-leagues"
      }
    }
  }
}

Then in Claude:

Set up my home league: 10 teams, full PPR, I pick 4th.
Set up my work league: 13 teams, half PPR, I pick 11th.
Switch to home. Sync my Sleeper draft 1234567890. Who should I pick?

Related MCP server: StatHead

Documentation

What it does

Recommends picks by opportunity cost, not raw value. Knowing a player is good is easy; knowing whether he'll still be there at your next turn is what decides who to take now. For each position the model walks the board top-down, accumulating the probability every better player is gone, and values a pick by its marginal gain over waiting.

At pick 6 with the top five off the board, that means recommending Amon-Ra St. Brown over Josh Allen — Allen grades as the highest-value player left, but has a 73% chance of lasting to pick 19, and the receiver has 8%.

Projects players from five seasons of open data. Offensive line from adjusted line yards and pressure allowed, neutral-script pace and run/pass split, five-year defensive strength by position, divisional schedule, injury history crossed with workload burden, positional aging curves. Computed from raw plays rather than scraped from ranking tables, so it recomputes under your scoring and doesn't break when a website changes its HTML.

Optimises for consistency. The default leans toward week-to-week reliability over raw ceiling. Tune it with one number.

Handles your actual leagues. PPR, half PPR, standard, superflex, TE premium, any size, any slot. Multiple leagues side by side with separate boards and drafts.

Separation and route efficiency from NGS tracking data — the open-data stand-in for paywalled charting. Validated against a published PFF table: TPRR within a few hundredths, ordering correct.

Rookies projected from draft capital, fitted to ten years of first-year outcomes.

Live sync from Sleeper (public API, no credentials), ESPN, or paste from anywhere.

Name matching that survives real drafts — JSN, CMC, Bijan, Hollywood Brown, Joshua Palmer, and typos. Ambiguous names name their candidates rather than guessing.

Tools

Data attribution

This project computes everything from open sources and ships no third-party data:

• nflverse — play-by-play, weekly stats, snap counts, injury reports, rosters, schedules, draft picks, combine.

• NFL Next Gen Stats (mirrored by nflverse) — separation, cushion, YAC over expected.

• dynastyprocess/data — FantasyPros expert consensus rank history.

Please respect the terms of those upstream projects. Nothing here scrapes paywalled sources, and no proprietary data is redistributed.