Chart Library
Integrations

Connect Chart Library to your agent.

Pick your path. The core research loop — search → pull_comps → cohort_introspect — is free and keyless on every one of them. Add an API key only when you outgrow 1,000 calls/day.

Fastest path · no install, no key

No framework at all? The flagship is one HTTP call. Paste this and run it right now:

curl -s https://chartlibrary.io/api/v1/cohort_analyze \
  -H "Content-Type: application/json" \
  -d '{"anchor":{"symbol":"NVDA","date":"2025-03-03","timeframe":"1d"},"horizons":[5]}'

Or try it live in the browser →

Hosted MCP connector (Claude) →
Keyless · no install
Add https://chartlibrary.io/mcp as a custom connector in Claude — OAuth, no API key, no config file. The full canonical toolset in about 30 seconds.
MCP server — Cursor & any MCP client →
MCP
Wire the canonical cohort tools into Claude Desktop, Cursor, or any MCP-aware client — via the hosted Streamable-HTTP endpoint or a local pip install.
OpenAI Agents SDK →
Python
Connect the MCP server with MCPServerStdio. The agent pulls the cohort of historical analogs instead of hallucinating technical analysis. Working code.
LangChain & LangGraph →
Python
Add the canonical cohort tools to a LangGraph ReAct agent in a few lines with langchain-mcp-adapters. Working code.
Vercel AI SDK →
TypeScript
A keyless tool() wrapper for cohort_analyze (~15 lines, fetch + zod), or connect the whole MCP surface. For Next.js / Node agents.
CrewAI →
Python
Give a crew calibrated history — connect the keyless MCP endpoint with MCPServerAdapter, or a native @tool. Keyless core loop.
Direct REST API + pip →
Any language
POST /api/v1/cohort_analyze from anything — keyless. Or pip install chartlibrary-mcp for stdio. Full endpoint and MCP-tool reference.
End-to-end: a Claude trading-research agent →
Walkthrough
The full cohort + memory loop architecture, start to finish, with working code — for when you want the whole pattern, not just the connection.

Whichever path you pick, you get the same engine: the cohort of historical analogs, a calibrated band, and the drivers — base rates, never a forecast. We measured the difference vs an ungrounded LLM: same coverage, a band 44% tighter, holding across Haiku, Sonnet, and Opus.