grahammccain/chart-library-mcp

What's new

Anthropic recommends roughly 8 tools per MCP server so the agent's tool-selection layer doesn't blur. v5 collapses the 19-tool sprawl into 8 canonical entry points plus a small write tool, with the remaining 12 kept as soft-deprecated wrappers for back-compat.

8 canonical tools (READ_ONLY)

• search — anchor a (symbol, date) and get the kNN cohort
• cohort — full conditional analysis (distribution, feature_importance, regime-stratified bands)
• discover — today's top setups across the market
• analyze — metric switchboard (anomaly, volume_profile, earnings_reaction, correlation_shift, pattern_degradation, regime_accuracy, …)
• context — market overview, ticker metadata, DB status
• narrative — realtime news pulse + alerts
• explain — prose narrative / filter importance / exit guidance / risk ranking
• portfolio — holdings analysis

Plus report_feedback (WRITE).

Tool annotations

Every read-only tool is tagged READ_ONLY so safe-mode clients (Cursor, Claude Desktop with strict permissions) auto-allow them. report_feedback is annotated WRITE.

Deprecation policy

The 12 legacy tools (search_charts, cohort_compare, live_search, compare_to_peers, decompose, detect_anomaly, cohort_analyze, narrative_pulse, narrative_alerts, discover_picks, symbol_intelligence, run_scenario) still work — they're thin wrappers that route to canonical tools and include a one-line hint in their description so agents migrate cleanly.

Why this matters for agents

Tool count is a real adoption-blocker — every additional tool dilutes the LLM's selection accuracy. With 8 high-leverage tools, agents pick the right one on the first try ~3× more often than with 19.

Install

pip install --upgrade chartlibrary-mcp

Configure

Same as before — add to your Claude Desktop config:

{
  "mcpServers": {
    "chartlibrary": {
      "command": "chartlibrary-mcp",
      "env": { "CHART_LIBRARY_API_KEY": "your_cl_key_here" }
    }
  }
}

Free Sandbox tier: 200 calls/day on search + context. Builder ($29/mo) for cohort, narrative, explain, deep endpoints.