MCP server for querying Roblox game portfolio data: recorded history via SQLite and live API readings
Five tools with clear, detailed descriptions (194 - 250 chars avg, well above 10-char minimum). All tools are read-only with appropriate naming (portfolio_now, live_now, game_history, list_games, game_analytics). Error handling is minimal, no recovery guidance, no actionable error messages. No tool annotations (readOnlyHint, etc.). Output schemas are not documented. Descriptions are excellent for context but lack LLM-optimized structure (e.g., 'When to use' vs 'When not to use').
Daily analytics for one game: revenue, DAU, retention, session length. metric: one of DailyActiveUsers, MonthlyActiveUsers, AverageSessionLengthMinutes, D1Retention, D7Retention, DailyRevenue, AverageRevenuePerUser, PayingUsers. since / until: optional plain YYYY-MM-DD DAY bounds, e.g. 2026-07-16. This is NOT the same format as game_history's since/until (a full ISO8601 timestamp): a bound in that shape is rejected here, loudly, not silently honoured. These figures trail real time by about two days, and different metrics complete at different times, so the response states last_complete_day. Do not read the newest row as today. For live player counts use live_now instead: these are daily aggregates, not a current reading.
Recorded time series for one game, oldest first. universe_id: the Roblox universe id (list_games has them). since / until: optional ISO8601 UTC TIMESTAMP bounds, e.g. 2026-07-16T09:00:00Z. This is NOT the same format as game_analytics' since/until (plain YYYY-MM-DD days): a bound in that shape is rejected here, loudly, not silently honoured. A null playing means that cycle did not measure it, which is not zero players.
Every universe the recorder has ever seen: newest name, first/last sample, sample count. Names change with events, so history is keyed on universe_id and the name shown is simply the most recent one recorded.
Call the Roblox API RIGHT NOW and return current values. Reads no history. Use for a spot reading at an arbitrary moment, at any resolution, without waiting for the 15 minute recorder cadence. Records nothing: a value read here does not enter the history.
Three tools (portfolio_now, live_now, list_games) have empty input schemas ({}). Even read-only tools with no parameters should declare an empty object schema explicitly with documentation.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract fields without knowing the response structure. game_history and game_analytics descriptions mention fields (e.g., 'age_seconds', 'last_complete_day') but no formal schema is visible.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All five tools are read-only and idempotent, marking them with readOnlyHint would signal safety to agents and enable optimizations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Latest RECORDED sample for every game, with age_seconds on each row. This is history, not this second. A row with a large age_seconds means the recorder has not sampled recently; use live_now for the current truth.
Error handling is minimal. query() function raises RuntimeError with generic messages ('query failed', 'recorder query returned not JSON'). No recovery guidance, no categorization (retryable vs fatal), no actionable next steps for LLMs.
game_analytics 'metric' parameter accepts free-form string instead of enum. Description lists valid values ('DailyActiveUsers, MonthlyActiveUsers, ...') but no enum constraint. LLMs may hallucinate invalid metrics.