MCP Server

MCP Server

The SDK ships a Model Context Protocol server that exposes CFTC COT data and analysis to MCP-compatible LLM clients (Claude Desktop, IDE assistants, etc.) over stdio transport. The server reuses COTClient and COTAnalysis internally — it doesn't call any LLM itself.

Installation

pip install cftc-cot-soda[mcp]   # requires Python >= 3.10

This registers the cftc-cot-mcp console script. Run it directly to test:

cftc-cot-mcp
# or: python -m cftc_cot.mcp_server

Claude Desktop config

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "cftc-cot": {
      "command": "cftc-cot-mcp"
    }
  }
}

Tools

Every market tool accepts dataset ("legacy" / "disaggregated" / "tff") and an exact flag (default false = case-insensitive prefix match). A market that matches nothing raises a clear error rather than returning an empty result.

Tool	Description
list_markets	List available markets for a dataset.
latest_report	Most recent COT report for a market.
history	N weeks of history (newest first).
net_positions	Net (long − short) per trader category.
cot_index	Classic 0–100 COT Index over a rolling window.
z_scores	Rolling Z-scores of net positions.
long_short_ratios	Long/short ratios per category.
wow_change	Week-over-week change in net positions.
percentile_rank	Percentile rank (0–1) of a net-position column's latest value.
extremes	Flag the latest bullish/bearish extreme.

Resources

URI	Description
cot://datasets	Dataset names and Socrata dataset ids.
cot://fields/{classification}	Field-name reference for legacy/disaggregated/tff.

Prompts

Prompt	Description
analyze_market(market, dataset)	Guide a full positioning analysis (COT Index → extremes → week-over-week).
positioning_summary(market, dataset)	Produce a concise positioning summary.

Notes

• The server uses an in-memory cache, so repeated queries within a session don't re-hit the CFTC API (COT data updates weekly).
• The mcp extra requires Python ≥ 3.10; the core SDK supports ≥ 3.9.