grahammccain/chart-library-mcp

Market memory for AI, with published research for agents and people.

Version 6.3.3 marks failed public tool calls as MCP errors while preserving their structured recovery details. Invalid requests explain how to correct the inputs; stale publication versions retain instructions to restart from the current overview. Successful response schemas and direct Python calls remain compatible.

Five read-only tools

Tool	Input	Result
market_state	Symbol; optional date	Completed-session state, historical analogs, outcome ranges, transition memory and tape
daily_note	Optional date	Published daily research, selection rule and settled-note tally
research_quality	None	Published calibration receipt, dated sample and qualifications
search_research	Query; optional kind, limit and offset	Relevant publications, IDs, findings, sample receipts, limitations and versions
read_research	Research ID; optional section, offset and version	Overview or an exact source document, with pagination and hashes

Market-state requests need no preliminary search or cohort handle. For published research, search for an ID, then read its evidence.

Example question: “Read AAPL's latest completed-session state. Report the historical analog ranges, sample sizes, session date and limitations.”

Omitting date uses the latest built/published session, not a real-time quote. Dates use YYYY-MM-DD. Missing or thin evidence stays missing or thin.

Connect

The primary remote endpoint is https://chartlibrary.io/mcp, using Streamable HTTP and no authentication.

For a command-based client, install or upgrade the package:

python -m pip install --upgrade chartlibrary-mcp

Then use the equivalent of this configuration in your MCP client:

{
  "mcpServers": {
    "chartlibrary": {
      "command": "chartlibrary-mcp"
    }
  }
}

No API key or server-side Python packages are required. An optional CHART_LIBRARY_API_KEY is forwarded as a Bearer token when set. An empty key does not send an Authorization header. CHART_LIBRARY_API_URL can select your own compatible endpoint; only point it at a server you trust with that key.

Python 3.10+ is required. This release uses the MCP 1.x FastMCP interface and pins mcp>=1.28.1,<2.0.0; MCP 2.x is a separate migration.

The equivalent HTTP calls

curl "https://chartlibrary.io/api/v1/state-packet?symbol=AAPL"
curl "https://chartlibrary.io/api/v1/daily"
curl "https://chartlibrary.io/api/v1/calibration"

Use only the call relevant to the question. For a historical state add &date=YYYY-MM-DD. Daily REST requests call the date parameter session.

Research workflow

search_research(query="IONQ noon", kind="casebook")
read_research(research_id="casebook:ionq-noon", section="article",
              version="FULL_VERSION_FROM_SEARCH")

Replace FULL_VERSION_FROM_SEARCH with the full 64-character version from the chosen search result. Updated servers supply documents[section].read_arguments directly; copy those into the tool and follow content.next_read until null. With earlier servers, pass the returned version and next_offset explicitly. Omitting the optional version reads the current publication without checking it against earlier responses. A changed version returns an error explaining how to restart from the current overview; the client does not retry automatically.

Search summaries are discovery metadata, not a source read. Read available guides and relevant limiting findings before making claims. Keyword search ranks document relevance, not market similarity or evidence strength. A partial result means a publication source was unavailable. Withdrawn articles cannot be read. Publication dates and market cutoffs differ; the noon IONQ case must not be merged with completed-session state analogs.

Complete IONQ example · Agent research plan

Read the evidence accurately

• Preserve dates, sample sizes, informative receipts, warnings and provenance.
• State-packet excess returns are percentage-point observations relative to a date-matched liquid-stock baseline. Each horizon has its own observed n.
• Raw historical percentiles are not automatically calibrated forecasts.
• research_quality audits only the method and population its receipt names. Do not transfer its coverage percentage to every market state or all research.
• Daily research has its own selection rule and settled-note tally.
• Historical frequencies are not recommendations to buy or sell.

Research access is free. Service limits and underlying data terms still apply. Honor HTTP 429 and retry guidance; do not assume unlimited throughput.

Existing integrations

Existing tool names stay registered and callable, including pull_comps, search, state_packet, and the cohort inspection tools. To discover the extended menu in a local client, set CHART_LIBRARY_MCP_PROFILE=advanced. This changes discovery, not authorization or access to private research.

The public menu does not expose private fund operations or every experimental state-memory method. Legacy examples in examples/ use the advanced interface.

Build with us

For a custom integration, a larger study, or a product built on the memory, contact Graham McCain — [email protected].

Developer guide · Methodology · Data terms · Privacy

Local verification

python -m pytest test_vendor_import.py

The import smoke runs without an API key and without application services/ or db/ packages. No package or registry publication is performed by tests.