Works with: Claude Desktop | Claude Code | ChatGPT | GitHub Copilot | Cursor | VS Code | Any MCP client
Cohort intelligence engine for stock chart patterns — give your AI agent the cohort of historical analogs, the full forward-return distribution, and the features that separated winners from losers. Calibrated, methodology-honest, no overstated confidence.
📖 What is cohort intelligence? · 🛠️ Full MCP setup guide · 🤖 Build an AI trading agent with Claude
25M+ pattern embeddings. 10 years of history. 19K+ stocks. One tool call.
> "What does NVDA's chart on 2024-08-05 1h look like historically?"
NVDA · 2024-08-05 · 1h — cohort of 500 historical analogs
(485 with realized 5-day returns)
Distribution at 5 days forward:
median: −1.3%
p10 ·· p90: −11.3% ·· +6.8% (80% empirical band)
win rate: 44%
cohort_score: 0.31 (modest)
Features that separated winners from losers:
+ credit_spread_state = tight
+ macro_state = bullish
+ pct_off_52w_low (further off)
− vol_regime = low
Summary: NVDA's 1-hour pattern on 2024-08-05 has 500 historical
analogs. The cohort's 5-day distribution is bearish-leaning
(median −1.3%, win rate 44%) — the historical record does NOT
show this pattern typically resolving bullish. Conditioning on
tight credit spreads and a bullish macro state would have
separated the outperformers within the cohort.
A retrieval, not a forecast. No hallucinated predictions. No cherry-picking. Just the empirical record your agent can cite.
pip install chartlibrary-mcpDownload the chart-library-6.1.0.mcpb extension file and open it with Claude Desktop for automatic installation.
claude mcp add chart-library -- chartlibrary-mcpAdd to claude_desktop_config.json:
{
"mcpServers": {
"chart-library": {
"command": "chartlibrary-mcp",
"env": {
"CHART_LIBRARY_API_KEY": "cl_your_key"
}
}
}
}Add to .cursor/mcp.json or VS Code MCP settings:
{
"servers": {
"chart-library": {
"command": "chartlibrary-mcp",
"env": {
"CHART_LIBRARY_API_KEY": "cl_your_key"
}
}
}
}Add to .vscode/mcp.json in your project (this file is already included in the chart-library repos):
{
"servers": {
"chart-library": {
"command": "chartlibrary-mcp",
"env": {
"CHART_LIBRARY_API_KEY": "cl_your_key"
}
}
}
}Copilot Chat will auto-detect the MCP server when you open the project. Use @mcp in Copilot Chat to invoke tools.
ChatGPT connects to MCP servers via remote HTTP endpoints. To set up:
- Enable Developer Mode: Go to ChatGPT Settings > Apps > Advanced settings > Developer mode (requires Pro, Plus, Business, Enterprise, or Education plan)
- Create a connector: In Settings > Connectors, click Create and enter:
- Name: Chart Library
- Description: Historical chart pattern search engine — 25M+ patterns across 19K+ stocks, 10 years of data
- URL:
https://chartlibrary.io/mcp - Authentication: No Authentication (or OAuth if using an API key)
- Use in conversations: Select "Developer mode" from the Plus menu, choose the Chart Library app, and ask questions like "What does NVDA's chart look like historically?"
Note: The remote endpoint at
https://chartlibrary.io/mcpuses Streamable HTTP transport. If you need SSE fallback, usehttps://chartlibrary.io/mcp/sse.
For any MCP client that supports remote HTTP connections:
https://chartlibrary.io/mcp
This endpoint supports both Streamable HTTP and SSE transports, no local installation required.
Free tier: 200 calls/day, no credit card required. Get an API key at chartlibrary.io/developers or use basic search without one.
> search(query="TSLA") → cohort_id
> explain(cohort_id=..., style="position_guidance")
Signal: HOLD
Of the historical analogs to this setup, those that exited early
avoided a drawdown 3/10 of the time; those that held gained a
further +2.1% median over the next 5 days. No exit signal triggered
— the cohort's record leans toward continuation, not reversal.
> context(target="market")
Sector relative strength (30-day):
Leaders: XLK Technology +4.2% · XLY Cons. Disc. +3.1% · XLC Comm. +2.8%
Laggards: XLU Utilities −1.4% · XLP Cons. Staples −2.1% · XLRE Real Estate −3.3%
Regime: Risk-On (growth > defensives), SPY above 20d, VIX mid-band.
> search(query="AMD 2024-06-18") → cohort_id
> cohort_groupby(cohort_id=..., by="ctx_spy_trend_20d")
AMD's cohort, split by the SPY trend at each analog's date:
SPY weak (bottom quartile): median 5d −5.2% · p10/p90 −11.4%/+1.1% · 18% positive
SPY strong (top quartile): median 5d +2.6% · p10/p90 −3.1%/+8.4% · 61% positive
A distribution conditioned on the tape — historical analogs, not a beta forecast.
Chart Library v6 exposes the same granular surface as the remote server at chartlibrary.io/mcp — so the pip package, the Claude connector, and the REST API all use the same tool names. The core loop is search → pull_comps → cohort_introspect. Chain tools via the comp_set_id / cohort_id handle for sub-second refinement without re-running kNN.
| Tool | What it does |
|---|---|
search |
Entry point. Find similar historical patterns for an anchor; returns a comp-set handle you can chain. mode= supports text (default), live_bars (raw OHLCV), similar (cohort-level neighbors). |
pull_comps |
The flagship. Pull the comp set for a subject (symbol, date, timeframe) — the historical analogs, what they did next, the drivers that separated the best outcomes, and our coverage_record. Front-of-house lexicon: subject · comp_set_id · comp_count · comp_strength · match_quality · drivers · up_rate · conditions (calm / normal / stressed). Same engine as cohort_analyze with the new vocabulary applied at the boundary. |
cohort_analyze |
Same engine as pull_comps under the original field names (cohort_id, feature_importance, win_rate, vol_regime, …). Kept callable verbatim for existing integrations; new ones should prefer pull_comps. |
cohort_introspect |
Slice/probe a stored comp set by ANY attribute (macro · technical · event) and get per-subset stats vs the full-cohort baseline. No kNN re-run. "Of the 300 analogs, how do the post-earnings-week ones do?" |
cohort_attribution |
Within-cohort winner/loser attribution — which member traits separated the forward-return tail from the rest, each with a by-date cluster-bootstrap CI and a false-discovery decision. Descriptive, never causal. |
track_record |
Historical predicted-vs-realized coverage of our calibrated bands (a track record, not a forecast). The nominal 80% band held 80.8% across 302,880 prior cases. |
symbol_intelligence |
Layer 5 memory — per-symbol feature reliability + achieved calibration across prior analyses. Ground a read in whether a feature has historically been reliable for this ticker. |
analyze |
Analytic metrics. metric= accepts anomaly, volume_profile, crowding, correlation_shift, earnings_reaction, pattern_degradation, regime_accuracy, decompose (slice winners vs losers), clusters (cohort-internal grouping). |
context |
Situational data. target= accepts "market", a ticker symbol ("NVDA"), {"symbol": ..., "date": ...} for lightweight anchor metadata, or "system" for DB coverage. |
explain |
Narrative + rankings derived from a cohort. style= accepts filter_ranking (which filter shifts the distribution most), prose (plain-English summary), position_guidance (exit signals), risk_ranking. |
portfolio |
Multi-holding weighted conditional distribution. Runs per-holding cohorts in parallel, weight-averages the distributions, ranks tail contributors. |
report_feedback |
File an error or improvement suggestion back to the project. |
Full-cohort handover — hand the raw cohort back so you can bucket/sort by your objective, not our default lens:
| Tool | What it does |
|---|---|
cohort_members |
The full cohort, one record per analog, with rich per-member metadata (forward outcomes, regime, anchor fundamentals, news, chart events). Slice and bucket it yourself. |
cohort_groupby |
Partition the cohort by one dimension (vol_regime, sector_etf, momentum_5d, …) → per-bucket outcome distributions vs baseline. The one-call "does this dimension matter?" primitive. |
cohort_rerank |
Reorder the cohort by a weighted composite of member fields you name (e.g. "ret_5d:1,distance:-0.5") — impose your objective on the analogs, fully auditable. |
These tools replace hallucinated "on average this pattern returns X%" with real conditional base rates. The full distinction — what they do and how to read responses — is documented at /concepts/cohort-intelligence and /concepts/reading-a-cohort-response.
1. search(query="NVDA 2024-06-18") → comp_set_id
2. pull_comps(symbol="NVDA", date="2024-06-18",
filters={"vol_regime": ["high"]})
→ comp set: distribution + drivers
3. cohort_introspect(cohort_id=...,
where={"events.days_since_earnings": {"max": 5}})
→ how the post-earnings subset did
4. cohort_groupby(cohort_id=..., by="sector_etf") → outcome split by sector
v6 converges on the granular naming the live remote/connector surface already used. The v5 umbrella tools — cohort (depth=), discover (mode=), narrative (mode=), and decision_brief — are now deprecated but still callable, so existing code keeps working. cohort(depth="full") forwards to cohort_analyze. New agents should reach for the canonical tools above.
| v5 umbrella call (deprecated) | v6 canonical |
|---|---|
cohort(depth="full", ...) |
cohort_analyze(...) |
cohort(depth="basic", cohort_id=...) then slice |
cohort_introspect(cohort_id=..., where={...}) |
cohort(depth="compare", compare_with={...}) |
cohort_compare(...) (still callable) |
portfolio(mode="symbol_intel", symbol=...) |
symbol_intelligence(symbol=...) |
| `discover(mode="picks" | "daily_setups")` |
| `narrative(mode="pulse" | "alerts")` |
The v4-era granular aliases (cohort_compare, decompose, clusters, live_search, similar_cohorts, anchor_fetch, narrative_pulse, narrative_alerts, discover_picks, get_daily_setups) remain deprecated-but-callable and forward to the canonical surface.
The v3-era tools (search_charts, get_cohort_distribution, analyze_pattern, etc.) were removed in v5. If your code still calls them, pin chartlibrary-mcp<5.0.0 until you migrate. The mapping:
| Legacy (removed in v5) | Replacement |
|---|---|
search_charts, search_batch, get_discover_picks |
search |
get_cohort_distribution, refine_cohort_with_filters, run_scenario, get_regime_win_rates, compare_to_peers |
cohort_analyze (+ cohort_introspect to refine) |
detect_anomaly, get_volume_profile, get_crowding, get_earnings_reaction, get_correlation_shift, get_pattern_degradation, get_regime_accuracy |
analyze (metric=) |
get_sector_rotation, get_status, get_market_context |
context |
get_pattern_summary, explain_cohort_filters, get_exit_signal, get_risk_adjusted_picks |
explain (style=) |
get_portfolio_health |
portfolio |
analyze_pattern, get_follow_through, check_ticker |
search + cohort_analyze |
Chart Library indexes a large library of historical chart patterns and exposes them behind a conditional-distribution API. Every query returns sample sizes, percentiles, and calibrated forward-return bands — never a point forecast.
When your agent calls search("NVDA") and chains cohort_analyze, the server:
- Resolves NVDA's current chart state to a stored embedding
- Retrieves the cohort of historically similar patterns
- Looks up what happened over the following 1, 3, 5, and 10 days
- Returns the calibrated distribution + a plain-English summary via Claude Haiku
The result: factual, citation-ready statements like "out of N similar historical patterns, the median 5-day return was X% (80% band [p10, p90])" that your agent can present without hallucinating or hedging.
| Tier | Calls/day | Price |
|---|---|---|
| Sandbox | 200 | Free |
| Builder | 5,000 | $29/mo |
| Scale | 50,000 | $99/mo |
Get your key at chartlibrary.io/developers.
export CHART_LIBRARY_API_KEY=cl_your_keyChart Library's privacy policy is published at chartlibrary.io/privacy and covers:
- What we collect: account info (email when you create an account), usage data (search queries, features used), and device information (browser, OS, IP). API queries are stored for service operation and analytics.
- How we use it: providing and improving the service, processing your searches, communicating about your account, and analyzing usage patterns.
- Data sharing: we do not sell personal data. Operational service providers (hosting, analytics, payment processing) receive only what's necessary to provide the service.
- Third-party services: queries may be processed by upstream providers (Polygon.io for market data, Anthropic for narrative summaries) under their own privacy policies.
- Retention: account info while your account is active; usage data is anonymized or deleted periodically. You can request deletion at any time.
- Security: encryption in transit and at rest. No method of transmission is 100% secure.
- California rights (CCPA): right to know, right to delete, right to opt-out, non-discrimination.
- Contact: support@chartlibrary.io for any privacy inquiry.
The MCP server itself sends only the arguments of your tool calls to chartlibrary.io (no local file or directory contents, no clipboard, no browser history). Your CHART_LIBRARY_API_KEY is sent only as a Bearer header to authenticate with the chart-library API.
- Transport: all calls to the remote API are HTTPS (TLS 1.2+).
- Authentication: optional API key passed as a Bearer header; the free Sandbox tier requires no key.
- No write access to your environment, files, or other accounts. The single MCP tool that performs a write (
report_feedback) only writes back to chart-library's own feedback inbox and never touches your system.
Report security issues to support@chartlibrary.io.
MIT. See LICENSE.
Chart Library provides historical pattern data for informational purposes. Not financial advice.