Guides
Working code, real data, calibrated outputs.
Practical guides for connecting your problem to Chart Library’s cohort intelligence primitive. Each guide is self-contained with runnable code.
Agent builders
MCP server for finance →
Wire Chart Library's 13 canonical cohort intelligence tools into Claude Desktop, Cursor, or any MCP-aware client. Install in 30 seconds.
Agent builders
Build an AI trading agent with Claude →
End-to-end walkthrough of the cohort + memory loop architecture that beat SPY +1.6pp in offline backtest. Working code.
Agent builders
Use Chart Library with LangChain & LangGraph →
Add the 13 canonical cohort tools to a LangGraph ReAct agent in a few lines with langchain-mcp-adapters. Working code.
Agent builders
Use Chart Library with the OpenAI Agents SDK →
Wire the cohort tools into an OpenAI Agents SDK agent with MCPServerStdio — real history instead of hallucinated TA.
Agent builders
Use Chart Library with the Vercel AI SDK (TypeScript) →
A keyless tool() wrapper for cohort_analyze in TypeScript — ~15 lines, no API key — or connect the whole MCP surface. For Next.js / Node agents.
Agent builders
Use Chart Library with CrewAI →
Give a CrewAI crew calibrated history — connect the keyless MCP endpoint with MCPServerAdapter, or a native @tool. No API key for the core loop.
Quants
k-NN regression for stock returns →
What works, what fails, and the calibration that makes it honest. From naive sklearn to production-grade cohort retrieval.
Engineers
Vector search for stock patterns with pgvector →
Production schema, IVFFlat tuning, embedding pipeline, and the operational gotchas at 25M embeddings scale.
Backtesters
Backtest a chart pattern honestly →
Survivorship bias and date-based splits inflate results 5-30 percentage points. Here's the discipline that prevents both.
Engineers
Trim the cohort_analyze response with fields= →
Drop up to 97% of response bytes when your backtest loop only needs the distribution. One parameter, full back-compat.