Best MCP servers for backtesting

Updated September 2026 · Written by QuantGPT, one of the six servers below, so read it knowing that

// The short answer

An MCP server is how an AI assistant such as Claude or ChatGPT reaches an outside tool. For backtesting, two questions decide it: where does the backtest run, and on what data?

If you want your AI to test an idea on decades of US stock history without writing code, QuantGPT runs the backtest on its own servers and sends back the numbers, losing years included. If you write code and want the same algorithm to go live, QuantConnect's open-source server drives its engine. BotSpot turns a plain-English idea into a Python bot for your broker. Panther tests indicator rules on crypto, forex, gold and silver, and Keel builds strategies for Hyperliquid perpetual futures. EODHD's server supplies global market data; its MCP pages describe no backtest tool.

Side by side
QuantGPT
EODHD
QuantConnect
BotSpot
Panther
Keel
Questions
How we compared

// Side by side

Facts about other servers come from their makers' own pages, checked 2026-09-27; each section below links its sources.
Server	Markets covered	History depth	What it does	AI clients	Price
QuantGPTthis site	US stocks and indexes; also crypto, commodities, FRED macro, prediction markets, sports odds and weather	S&P 500 index levels from 1928; single stocks thin before 1980; delisted companies priced from 1998 but not yet in market-wide backtests	Runs backtests, parameter sweeps, screens and event studies on its own servers; read-only SQL over its warehouse; no order tools	Claude (claude.ai, Desktop, Code), ChatGPT, Cursor, Gemini, any MCP client with a key	Free: 5 backtests and 5 AI calls, no card. QuantGPT $49 a month, QuantGPT Pro $149 a month
EODHD MCPdata	70+ exchanges: stocks, ETFs, indices, forex, US options; fundamentals, news, macro, ESG (per its MCP page)	30+ years on paid plans; the past year on the free plan	72 read-only data tools; technical indicators and pivot levels computed server-side; no backtest tool described	Claude, Claude Code, ChatGPT, Cursor, Windsurf, custom MCP clients	Free: 20 API calls a day. Personal plans $19.99 to $99.99 a month
QuantConnect MCPopen source	Equities, indexes, forex, crypto, futures and options	US equities from 1998, tick to daily	60+ tools over its API: write and compile code, backtest, optimize, deploy live. Runs on your machine	Claude Code, Claude Desktop, Copilot, Cursor, Devin Desktop	Uses API access, which starts at the paid Researcher tier; no flat price listed
BotSpot MCPby Lumiwealth	Stocks, options, crypto, futures	Not published on its MCP pages	32 tools: plain English to a Python bot, hosted backtests, charts, deploys to 10+ brokers	Claude, ChatGPT, Claude Code, Cursor, Codex	Free start; paid plans $49 to $399 a month at list price
Panther MCPby PantherWatch	30+ assets: crypto pairs, forex pairs, gold and silver	Not published in its README	13 tools: plain-English indicator backtests, parameter optimization, portfolio backtests, share links	Claude Desktop, Claude Code, ChatGPT (Pro, Team, Enterprise), Gemini CLI, Cursor, Cline, VS Code Copilot	Free API key at sign-up; no prices in its README
Keel MCPHyperliquid	Hyperliquid perpetual futures, about 220 of them	BTC, ETH and SOL from June 2024; HYPE from December 2024	Deterministic backtests with fees, funding and slippage; opt-in live trading on Hyperliquid	Claude (web, desktop, phone), ChatGPT, Claude Code, Cursor, Codex, Windsurf	Free: 50 backtests a week. $29 to $199 a month, plus a trading fee of 1 to 5 bps

// The six servers, one by one

QuantGPTterminal.quantgpt.co/mcp · this site

Best forAsking research questions of decades of US market history from Claude or ChatGPT, with the backtest computed on the server and the losing years shown.

ConnectThe address is https://terminal.quantgpt.co/mcp (Streamable HTTP). Claude and ChatGPT sign in over OAuth with no key to paste; Claude Code, Cursor, Gemini CLI and scripts send a key from your account page. Guides: Claude, ChatGPT, Gemini, Cursor.
DoesThe backtest runs on QuantGPT's servers and the result comes back with the numbers behind it, losing years included. Tools cover rule backtests, parameter sweeps, screens, stock grades, event studies and read-only SQL across the warehouse.
Library175 documented strategies, 63 of them runnable in one call from your AI; the rest are read-only for now.
LimitsCompanies that later delisted are priced from 1998 but are not in the market-wide backtest universe yet, so a market-wide backtest is not survivorship-free; a backtest on names you list, or a rule on a single index, is not affected. Fundamentals are joined by filing date, but about 9% of rows still carry the quarter-end as the filing date (a fix is in progress). Stock coverage is the US market.
Check itThe public verify page shows our daily bars one at a time, to hold against any independent source. No tool trades: nothing your AI calls places, changes or cancels an order. The hosted runner, your own code at your own broker, is coming to QuantGPT Pro.

Plans: Free is 5 backtests and 5 AI calls for the life of the account, no card. QuantGPT is $49 a month (200 backtests and 500 AI calls a day, 3 keys) and QuantGPT Pro is $149 a month (1,500 AI calls a day, 5 keys), each with 3 days free (5 backtests and 5 AI calls) and a card up front. Compare the plans or start free.

Pick another server if you need listings outside the US, orders sent to a broker today, or more free use than 5 backtests and 5 AI calls.

EODHD MCPmcp.eodhd.com · market data

Best forLooking up global market data inside your AI: prices, fundamentals, news, options and macro, with a free tier to start.

Tools72 read-only tools in 15 categories, across 70+ exchanges, with 30+ years of history.
ConnectTwo endpoints: v1 takes an API key in the URL (its page names ChatGPT, Claude Code and custom agents) and v2 uses OAuth 2.1 (Claude Desktop).
DoesFetches data. Its docs list server-side technical indicators and support and resistance levels by five pivot methods; its MCP pages describe no backtest tool, so a backtest on its data is up to your AI or your own code.
PriceFree plan: 20 API calls a day and the past year of data. Personal-use plans run $19.99 to $99.99 a month with 30+ years of history; commercial use is priced separately. MCP tool calls use the same quota as regular API requests.

Sources, checked 2026-09-27: eodhd.com/lp/eodhd-mcp-server, eodhd.com MCP docs, eodhd.com/pricing. The longer version: QuantGPT vs the EODHD MCP server.

QuantConnect MCPopen source · runs on your machine

Best forDevelopers who want an AI to write, backtest and deploy full algorithms on QuantConnect's engine, across many asset classes.

Tools60+ tools covering the QuantConnect API: projects, files, compiling, backtests, parameter optimization and live trading. Fully open source.
ConnectRuns locally, in Docker with your QuantConnect user ID and API token, or inside VS Code, which its README calls the preferred way (the README marks the standalone repository deprecated). Its docs name Claude Code, Copilot, Cursor and Devin Desktop; the README walks through Claude Desktop.
DataUS equities from 1998, tick to daily, a dataset QuantConnect's docs describe as survivorship-bias-free. The free plan's data list covers equities, indexes, forex, crypto, futures and options.
PriceThe Free plan lists unlimited backtesting but no API access. API access, which the MCP server uses, comes with the paid tiers from Researcher up, priced by the setup you choose.

Sources, checked 2026-09-27: quantconnect.com/mcp, MCP server docs, GitHub README, US Equities dataset, quantconnect.com/pricing.

BotSpot MCPmcp.botspot.trade/mcp · by Lumiwealth

Best forTurning a plain-English idea into a Python trading bot, backtesting it, and running it at your own broker.

Tools32 tools in 8 groups: strategy generation, strategy management, backtesting, deployment, visuals, marketplace, account and data. Strategies are Python on the open-source Lumibot library.
ConnectOAuth 2.1 for Claude and ChatGPT, or an API key for Claude Code, Cursor and Codex. Its site also names Gemini.
MarketsStocks, options, crypto and futures, deployable to 10+ brokers including Charles Schwab, Interactive Brokers and Alpaca. The history depth of its backtest data is not published on its MCP pages.
PriceA free start (its README lists 10 AI messages, 2 strategy generations and 30 minutes of backtesting a month). Light, Starter, Pro and All Access list at $49, $99, $199 and $399 a month; on 2026-09-27 its pricing page showed Light and Starter at 50% off.

Sources, checked 2026-09-27: GitHub README, Lumibot docs: BotSpot MCP, botspot.trade, botspot.trade/pricing.

Panther MCPpanther.watch/mcp · by PantherWatch

Best forQuick plain-English backtests of indicator rules on crypto, forex, gold and silver.

Tools13 tools: list assets, fetch prices, run and poll backtests, optimize parameters, run weighted portfolio backtests, and get a shareable link to each result.
Markets30+ assets: 20 crypto pairs via Binance, 10 forex pairs via IC Markets, plus gold and silver. Its README lists no stocks.
Rules25 indicators, including SMA, EMA, RSI, MACD and Bollinger Bands; long, short or both directions.
ConnectThe remote endpoint, or locally with uvx panther-mcp and an API key. Clients in its README: Claude Desktop, Claude Code, ChatGPT (Pro, Team and Enterprise in developer mode), Gemini CLI, Cline, Cursor and VS Code Copilot. The key is free at sign-up; the README gives no prices and no history depth.

Source, checked 2026-09-27: GitHub README.

Keel MCPmcp.usekeel.io/mcp · Hyperliquid perps

Best forStrategies on Hyperliquid perpetual futures, where the same compiled strategy runs in the backtest and live.

DoesComposes strategies from typed components and runs deterministic backtests with real fees, funding and slippage across about 220 perps; its README also offers walk-forward and Monte Carlo checks.
DataBTC, ETH and SOL price history starts in June 2024 (carry from late July 2024), HYPE on December 5, 2024, and order-flow series in March 2025.
ConnectA hosted endpoint, or a local command-line tool and server (pipx install keel-trade). Clients named: Claude on web, desktop and phone, ChatGPT, Claude Code, Cursor, Codex and Windsurf. Live trading is opt-in: the live tools need an explicit toolset and a local arming step.
PriceFree: 50 backtests and 20 AI messages a week, 1 live strategy, and a 5 bps trading fee. Starter $29, Trader $79 and Pro $199 a month, with the trading fee down to 1 bp on Pro.

Sources, checked 2026-09-27: usekeel.io/keel-mcp, GitHub README, usekeel.io/pricing.

// Questions

What is the best MCP server for stock market backtesting?

It depends on who does the work. As of September 2026: if you want your AI to test a rule on decades of US stock history without writing code, QuantGPT runs the backtest on its own servers and sends back the numbers, losing years included; its market-wide backtests do not yet include companies that later delisted. If you write code and want the same algorithm to trade live, QuantConnect's open-source MCP server drives its engine, with US equity data from 1998. BotSpot turns a plain-English idea into a Python bot that runs at your broker. Panther (crypto, forex, gold and silver) and Keel (Hyperliquid perpetual futures) serve other markets. EODHD's server supplies market data, and its MCP pages describe no backtest tool. QuantGPT publishes this page, so weigh that.

How do I connect Claude to real stock market data for backtesting?

Add an MCP server as a custom connector. In claude.ai or Claude Desktop, open Customize, then Connectors, click + and choose Add custom connector, paste the server's URL and sign in when Claude asks. For QuantGPT the URL is https://terminal.quantgpt.co/mcp and the sign-in needs no key; in Claude Code, run claude mcp add with a key from your QuantGPT account page. Then ask Claude to run a backtest and it calls the server's tools mid-conversation. EODHD's data server connects the same way; its page gives an OAuth endpoint for Claude Desktop and a key-based one for other clients. Step by step: quantgpt.co/learn/connect-claude.html.

Can AI assistants run real backtests on historical market data?

Yes, when the MCP server they connect to has a backtest engine: the AI sends the rules, the server computes them on its stored history, and the results come back into the chat. As of September 2026, QuantGPT, QuantConnect, BotSpot, Panther and Keel all describe backtest tools on their MCP servers. A data-only server returns rows, and the AI then has to do the math inside its own context window, on whatever fits. Either way a backtest is hypothetical: check the sample size, the losing years, the trading costs, and whether companies that later failed are in the data.

How can I backtest a trading strategy with ChatGPT?

Connect ChatGPT to an MCP server that runs backtests, then describe the strategy in plain English. In ChatGPT, open Settings, then Security and login, and turn on Developer mode (OpenAI offers it on Plus, Pro, Business, Enterprise and Education accounts, on the web). Then go to chatgpt.com/plugins, select the plus button, name the app and enter the server's URL under Connection. For QuantGPT that is https://terminal.quantgpt.co/mcp, and you sign in to QuantGPT with no key to paste; on a Business, Enterprise or Education workspace an admin may need to allow developer mode first. Panther's README lists ChatGPT Pro, Team and Enterprise in developer mode, and BotSpot and Keel list ChatGPT too. Step by step: quantgpt.co/learn/connect-chatgpt. Without developer mode, QuantGPT's paste-pack gives ChatGPT the context in one block, but ChatGPT then cannot run the backtests itself.

// How we compared

QuantGPT publishes this page, and QuantGPT is one of the six servers in it. Every fact about another server comes from its maker's own public pages, GitHub README or docs, linked in its section, as read in September 2026. We did not run the other servers or time them. Tool counts, data and prices change often, so check the linked page before you pay. Facts about QuantGPT come from its own plans, guides and product.

This is not every MCP server that touches markets. We picked servers whose makers publish enough to compare on markets, history, tools, clients and price. Where a maker does not publish a fact, the table says so rather than guess. If you build one of these servers and something here is wrong or out of date, write to desk@quantgpt.co and we will correct it.

Facts about other companies are from their public pages as of the date above and may have changed. EODHD, QuantConnect, BotSpot, Panther, Keel and every other product named here are trademarks of their owners; no affiliation or endorsement is implied. Backtests are hypothetical and gross of costs unless stated, and past results, real or simulated, do not predict future results. QuantGPT is research software, not investment advice, and never places a trade. The terms govern.