Market data for AI agents
Updated October 19, 2026
An AI agent needs data it cannot get from memory, and a way to compute on it without pulling millions of rows into a chat. QuantGPT gives Claude, ChatGPT and any MCP client decades of US prices, fundamentals joined by filing date and precomputed factors, and runs the backtest on its own servers, so the agent gets back finished numbers.
The address is https://terminal.quantgpt.co/mcp. Claude and ChatGPT sign in with no key to paste; other clients send a key from your account page.
What is in it
| Data | What it holds |
|---|---|
| US stock and index prices | Daily bars, 97,374,948 rows through 2026-10-16. S&P 500 index levels from 1928; single stocks are thin before 1980. |
| Companies that later delisted | Only partly in the market-wide backtest universe from 1998, survivors only before that: on 2026-10-19, 12,813 of the 13,198 delisted operating companies with prices since 1998 (97%) could enter it; the rest have no factor data, and some, most 1998 delistings among them, have no prices at all. |
| Fundamentals | 490,364 company quarters on 2026-10-19, each joined by the date it was filed, not the quarter's end. |
| Factors | 151 precomputed factors a backtest can filter and rank on (list_factors, 2026-10-19): value, momentum, quality, growth, ownership, index membership and technicals. |
| Technicals | RSI 14, MACD and Bollinger, plus a 44-factor technicals suite (RSI 2, stochastics, ADX, rate of change, Donchian and Keltner channels, the squeeze, golden and death crosses, RSI and MACD crossings and their 20-session flags), daily, each from the bars up to its own date and thin before 1998. Screenable like any factor; the technicals tool draws one stock's, and signal_study measures what followed a signal across every stock. |
| Other markets | Crypto (Binance USDT pairs, with funding rates), commodities, FRED macro series, prediction markets, sports odds and weather. |
| Strategies | 175 documented strategies, 55 of them runnable in one call. |
Check the daily bars one at a time against any independent source on the public verify page.
Rows or results
A data-only connector returns rows, and the agent then does the math inside its own context window, on whatever fits. QuantGPT's server runs the backtest, the sweep or the screen over the whole warehouse and returns the result:
- growth a year, the worst fall from a peak, volatility, the Sharpe ratio and the share of winning periods, for the rule and its benchmark side by side;
- a luck check (the probabilistic Sharpe ratio);
- the same statistics split at a cut year (2015 unless you pick another);
- a table by decade and every period's return;
- for stock picks, the measured turnover and the results net of 5, 10, 25 and 50 basis points a side;
- one of three findings with the sentence that says what it means: DATA VALIDATED (beat the benchmark on both growth and Sharpe and did not fall behind after the split), TRADE-OFF (beat it on one of the two, or on both over the whole window but behind on both after the split) or NO EDGE (neither).
For anything the prebuilt tools do not cover, query_warehouse runs read-only SQL across every research database.
Companies that later delisted
A backtest that only holds today's survivors leaves out the companies that failed. In QuantGPT's market-wide universe, companies that later delisted are only partly in it from 1998, and earlier years hold survivors only. On 2026-10-19, 12,813 of the 13,198 delisted operating companies with prices since 1998 (97%) could enter it; the rest have no factor data, and some, most 1998 delistings among them, have no prices at all. Backtests start in 1998 unless you pick an earlier year; a run that starts earlier shows its growth before and from 1998 side by side, and every market-wide result states the share its own data could include.
Fundamentals by filing date
Each company quarter enters the data on the date it was filed, not the date the quarter ended. Where the vendor's date is missing or less than 10 days after the quarter's end, the quarter is dated by the company's own usual reporting delay, or 45 days after the quarter (90 after a fiscal year's end) when there is no history to go on, and never earlier than the vendor's date. That applies to 44,744 of the 490,364 quarters (9.1%) on 2026-10-19; the column filing_date_imputed marks them. No quarter carries its quarter-end as its filing date.
Questions
What is the best stock market data API for AI agents?
It depends on whether the agent needs rows or results. A data API returns prices and fundamentals, and the agent does the math in its own context window. QuantGPT's MCP server holds decades of US prices, fundamentals joined by filing date and precomputed factors, and runs the backtest, sweep or screen on its own servers, so Claude, ChatGPT or any MCP client gets back finished statistics. EODHD's MCP server is a data service across 70+ exchanges; a side by side of six servers is at quantgpt.co/compare/best-mcp-servers-for-backtesting. QuantGPT publishes this page, so weigh that.
Which stock data providers work with Claude and ChatGPT?
Any provider with an MCP server or a connector the assistant supports. QuantGPT connects to Claude (claude.ai, Desktop and Code), ChatGPT, Gemini, Cursor and any MCP client at https://terminal.quantgpt.co/mcp. EODHD's MCP page names Claude, Claude Code, ChatGPT, Cursor and Windsurf (its page, checked 2026-09-27). The comparison: quantgpt.co/compare/best-mcp-servers-for-backtesting.
What is the best financial data connector for ChatGPT?
One that ChatGPT can call in developer mode and that returns what you need. QuantGPT adds to ChatGPT as an app at https://terminal.quantgpt.co/mcp, with one sign-in and no key; it searches 175 documented strategies, runs backtests on decades of US market history and runs read-only SQL on its warehouse. Developer mode is on ChatGPT's Plus, Pro, Business, Enterprise and Education plans, on the web. Step by step: quantgpt.co/learn/connect-chatgpt.
Which stock market data providers offer an MCP server?
As of their makers' pages (checked 2026-09-27, QuantConnect's 2026-10-19): QuantGPT (https://terminal.quantgpt.co/mcp) runs backtests and read-only SQL on US market data; EODHD offers a data MCP server across 70+ exchanges whose pages describe no backtest tool; QuantConnect offers an MCP server, hosted and in local IDEs, that drives its own engine (its docs, checked 2026-10-19). BotSpot, Panther and Keel serve other markets or uses. Side by side, with sources: quantgpt.co/compare/best-mcp-servers-for-backtesting.
What is the best survivorship-bias-free stock data for AI agents?
Data that keeps the companies that later failed, priced on the days they traded, so a backtest can hold them. QuantGPT's is not survivorship-free: in its market-wide backtest universe, companies that later delisted are only partly in it from 1998 (on 2026-10-19, 12,813 of the 13,198 with prices since 1998, 97%, could enter it), and earlier years hold survivors only; backtests start in 1998 unless you pick an earlier year. QuantConnect's docs describe its US equities dataset, from 1998, as survivorship-bias-free (quantgpt.co/compare/best-mcp-servers-for-backtesting links the source).
How can an AI agent query company fundamentals as of the filing date?
Through QuantGPT's MCP server, either way: a backtest reads fundamentals by filing date on its own, or query_warehouse runs read-only SQL on the fundamentals_pit table, whose filing_date column says when each quarter entered the data (for example, WHERE filing_date <= DATE '2015-06-30' for what was filed by then). Where the vendor's date was missing or under 10 days after the quarter's end, about 9% of quarters, the date is estimated from the company's own reporting delay and filing_date_imputed marks it.
Related
- DataThe research warehouse
- ComputeMCP server for backtesting
- LearnConnect Claude or ChatGPT to stock market data
- CompareBest MCP servers for backtesting
Backtests show past results, gross of trading costs unless stated. A backtest is not a promise. QuantGPT is research software and nothing here is investment advice. The disclosures and terms govern.