For Claude users + agent builders

Use Chart Library inside Claude.

Chart Library is an MCP server. Connect it to Claude and every question about a stock pattern is answered with historical analogs, forward-return distributions, and calibrated bands — grounded retrieval, not a generated prediction. 25M+ patterns spanning 10 years of US equity history, drawn from a 19,000+ symbol universe. One tool call.

Three ways to install

1. Claude Desktop — one-click .mcpb install

easiest

Download the desktop extension bundle and open it with Claude Desktop. Adds the 8 canonical tools (search, cohort, discover, analyze, context, narrative, explain, portfolio) to your local Claude. Optional API key for higher volume — the free Sandbox tier covers the full core loop, including the flagship Layer 3 cohort intelligence, without one.

Download the latest .mcpb release →

2. Remote MCP — add to Claude as a custom connector

one-click for paid plans

Pro / Max / Team / Enterprise users can wire the hosted MCP endpoint directly: Settings → Connectors → Add custom connector, then paste:

Name:        Chart Library
MCP URL:     https://chartlibrary.io/mcp
Transport:   Streamable HTTP (auto-detected)
Auth:        OAuth 2.0 (recommended) or API key Bearer header

Claude.ai bounces you to chartlibrary.io for the OAuth consent screen — sign in (or create a free account), click Approve, and you're back in Claude with the connector active. No API key to copy, no config file to edit. The flow uses OAuth 2.0 authorization- code with PKCE (RFC 6749 + RFC 7636), discoverable via /.well-known/oauth-authorization-server.

The endpoint speaks the MCP 2025-03-26 spec. Origin-header validation is enforced per the spec; the standard Anthropic surfaces are allowlisted. Free Sandbox tier (1,000 calls/day) is included with any account.

3. Claude Code / Python SDK — install from PyPI

for agent builders

For Claude Code, Python agents, LangChain, CrewAI, OpenAI Agents, or any framework with MCP support. Stdio transport — runs locally, calls our HTTP API under the hood.

pip install chartlibrary-mcp

# Claude Code:
claude mcp add chart-library -- chartlibrary-mcp

# Claude Desktop (manual config):
# Add to claude_desktop_config.json:
{
  "mcpServers": {
    "chart-library": {
      "command": "chartlibrary-mcp",
      "env": { "CHART_LIBRARY_API_KEY": "cl_your_key" }
    }
  }
}

What it does

Eight canonical tools. Ask Claude any question about a stock pattern and it'll reach for these instead of hallucinating from training data.

search	Find similar historical patterns, return a cohort_id you can chain
cohort	Conditional distribution — p10/p25/p50/p75/p90 + calibrated bands + feature importance + regime stratification
discover	What's interesting today — picks, daily setups, risk-adjusted ranking
analyze	Anomaly, volume, crowding, earnings reaction, regime accuracy, pattern degradation
context	Market state, ticker metadata, anchor regime context
narrative	News intelligence — single-symbol pulse or market-wide anomalies
explain	Prose narrative, filter importance, exit guidance, risk ranking
portfolio	Multi-holding cohort, or per-symbol track record

Example session

User:   What does NVDA's chart on 2024-08-05 1h look like historically?

Claude: I'll search Chart Library for analogs.
        → cohort(symbol="NVDA", date="2024-08-05", timeframe="1h", depth="full")

        Found 293 historical analogs. 5-day forward distribution:
          median:   −0.81%
          p10/p90:  −12.15% / +10.71%   (80% conformal band, validated 80.8%)
          win rate: 47%

        Volatility regime is the key signal: in high-vol setups like
        today's (vix=18, vix_term_structure=0.92), this pattern hit
        70% win rate with mean +2.73%. In low-vol it hit 26% / −4.45%.

        Today's setup matches the favorable high-vol profile, but
        magnitude is wide. Position size accordingly.

        Source: Chart Library cohort_analyze, n=293, validated
        calibration. Not a forecast.

Why use it

Grounded retrieval, not generation. Every response cites a specific cohort. Sample size, percentiles, and calibrated bands are baseline output.
Calibration is the moat. Our nominal 80% empirical band has held 80.8% across 310,592 audited cases. Validate against /calibration or call /api/v1/calibration directly.
Compliance-safe framing. Output is a fact about historical outcomes, not a forecast. Designed to be quoted by an agent without triggering "investment advice" classifiers.
Composable. Chain tools via cohort_id handles. Refine filters without re-running kNN.

Pricing

Tier	Calls included	Layer 3	Price
Free	1,000/day	✓	Free
Builder	50K/mo	✓	$29/mo
Scale	500K/mo	✓	$99/mo
Agent	2M/mo	✓	$299/mo
Enterprise	Unlimited	✓	From $2K/mo

Get a key at /developers. The full core loop — search, context, explain, and the flagship Layer 3 cohort intelligence (pull_comps / cohort_analyze) — is free, keyless over the hosted MCP or with a free key (1,000 calls/day). Paid tiers add volume, higher RPM, and SLA — you convert at scale, not to unlock the product.

Trust + compliance

Privacy — /privacy covers collection, usage, sharing, retention, contact. Tool calls send only the arguments you pass — no file system, no clipboard, no browser history.
Transport — HTTPS (TLS 1.2+) end-to-end. Origin-header validation per the MCP spec.
Write surface — exactly one write tool (report_feedback), which writes back to our own feedback inbox. Never touches your system or other accounts.
Disclaimer — /disclaimer. Chart Library is historical data analysis, not investment advice.

/learn/intelligence — what cohort intelligence is, methodology in plain English
/learn/methodology — how the cohort + calibration math works
/calibration — live calibration report (80% band coverage tracked)
/guides/mcp-server-for-finance — full MCP setup guide with worked examples
/guides/build-ai-trading-agent-claude — end-to-end agent template using Chart Library
github.com/grahammccain/chart-library-mcp — source, releases, issue tracker