NBA player intelligence workbench. Project next-game stats, compare against live lines, analyse props, score hypothetical scenarios, and generate AI analytical prose — all grounded in Hooplytics' RACE ML models and The Odds API.
Hooplytics MCP server has well-structured tool definitions with clear descriptions and mostly complete schemas. All 9 tools have descriptions (194 - 400 chars, within baseline 34 - 392 range). Input schemas are present and typed. However, output schemas are not documented, LLMs cannot predict return structure. Parameter descriptions are good but some lack constraints (e.g., 'last_n' has no min/max bounds). Error handling is minimal; no recovery guidance. Tool naming is mostly clear (project_player, analyze_prop, fantasy_props) but 'manage_roster' conflates view/add/remove into one tool. The 'features' parameter across multiple tools is underdocumented, callers must manually construct rolling feature dicts, forcing external knowledge. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Compare a player's model projection against a sportsbook line for a single stat. Returns the projected value, MORE/LESS call, edge gap, historical hit-rate, and model confidence metadata.
Retro-evaluate the RACE models against historical actuals for a player/stat. Returns a table of past games with projections, actuals, errors, and directional accuracy. Requires cached historical odds (from ingest-odds) for evaluation.
Run the Hooplytics Role Shift Detector for a player using their current rolling features. Checks 4 signals (assists, points, usage, minutes) against season baselines and flags in-series role transitions that would break RACE directional calls.
Generate a full 8-stat decision table for a player across all RACE models. Returns MORE/LESS calls for points, rebounds, assists, PRA, 3PM, stl+blk, turnovers, and fantasy score with edge metrics and confidence.
View, add, or remove players from the shared CLI/MCP training roster. The roster is persisted to ~/.hooplytics/roster.json and used by both the CLI and MCP server to determine which players' game logs are fetched and which are included in model training.
Output schemas not documented. LLMs cannot predict return structure (dict keys, field types, nested objects). Baseline: 100% of A+ tools document return types.
manage_roster conflates three distinct actions (view/add/remove) into one tool. Should split into list_roster, add_to_roster, remove_from_roster for single responsibility.
Numeric parameters lack bounds. 'last_n' (default 5) has no min/max; 'max_tokens' (default 800) unbounded. LLMs may pass absurd values (last_n=10000, max_tokens=1000000).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Project a player's next-game stats across all 8 Hooplytics RACE models (points, rebounds, assists, PRA, 3PM, stl+blk, turnovers, fantasy score). Returns a dict keyed by stat name with each model's projection, target column, and feature inputs. If features are provided, the Role Shift Detector runs and annotates suppressed stats in the response.
Score a hypothetical box-score row (e.g. 'what if LeBron had 28 points, 8 rebounds, 6 assists?'). Returns fantasy score and per-stat predictions from all 8 RACE models.
Generate AI analytical prose for a player using Claude or GPT. Synthesizes projections, live lines, role-shift context, and historical performance into a narrative brief.
Generate a brief AI summary of the day's slate (all games + player matchups). Synthesizes projections, live lines, and role-shift context into a concise narrative.
'features' parameter across 4 tools is underdocumented. Callers must manually construct rolling feature dicts (ast_l3, ast_l30, pts_l3, etc.). No helper tool to compute these; forces external knowledge.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). manage_roster has WRITE risk but no annotation. Agents cannot distinguish safe reads from state-modifying calls.