MCP server exposing normalized CME futures data: back-adjusted continuous bars, contract specs and summary stats.
Five tools with clear verb-noun naming (ping, list_instruments, get_ohlcv, get_contract_specs, get_summary_stats). All have descriptions (avg 180 chars, within 10-1024 baseline). Input schemas present with type definitions and parameter descriptions. Tool annotations correctly applied (readOnlyHint=true, idempotentHint=true). Error handling via ToolError with domain-specific exceptions (InstrumentError, TimeframeError, ValueError) mapped to user-facing messages. Key gaps: output schemas not documented in code; no pagination guidance for list_instruments despite potentially large result sets; parameter constraints (e.g., timeframe enum values) stated in descriptions but not formalized in schema; no recovery guidance in error messages.
Return contract specifications for a registered futures symbol. Includes tick_size, point_value, tick_value (USD/tick), exchange, commission_per_side, initial_margin_day, currency and RTH session hours (ET).
Return OHLCV bars for a futures instrument. Daily requests for a back-adjusted symbol are served from 'databento-badj' (adjusted=true); everything else from 'yahoo-unadjusted' (adjusted=false). The response always labels 'source' and 'adjusted'. Filter with ISO dates (start/end) or n_bars (most recent N). Output is capped (most recent rows); check 'truncated' and 'n_total'.
Summary statistics for a series without returning the bars. Returns period return, annualized volatility (approximate), max/min close, max drawdown and bar count. Cheaper than get_ohlcv when you only need a high-level read. Labels 'source' and 'adjusted'.
List available futures instruments. For each symbol reports whether contract specs exist and, per data source, the timeframes and date range. Sources: 'databento-badj' (back-adjusted continuous, DAILY only) and 'yahoo-unadjusted' (stitched front-month, NOT adjusted). Use this first to discover what to request.
Health check. Returns 'pong' if the server is wired correctly.
Output schemas not documented. Tool descriptions state what fields are returned (e.g., 'source', 'adjusted', 'truncated', 'n_total' for get_ohlcv) but no formal schema definition visible in code. LLMs cannot plan downstream calls or extract fields reliably without documented return types.
Parameter constraints stated informally in descriptions, not formalized in schema. E.g., timeframe 'One of 1m, 5m, 15m, 1h, 1d' should be an enum; date format 'ISO date/datetime UTC' should be a pattern. Formal constraints prevent LLM hallucination and enable validation.
list_instruments returns potentially large result set with no pagination, limit, or offset parameters. Description does not state result cap or guidance on when to call. Large unfiltered lists risk context window exhaustion.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Error messages lack recovery guidance. ToolError is raised with domain error text (e.g., 'InstrumentError'), but no hint on next steps (e.g., 'Call list_instruments() to see available symbols'). Agents cannot self-correct without explicit guidance.
get_ohlcv parameter 'n_bars' lacks bounds. Description does not specify min/max (e.g., 1 - 10000). Unbounded integers let LLMs pass absurd values that could timeout or exhaust memory.