MCP server for backtesting
Updated October 19, 2026
QuantGPT's MCP server at https://terminal.quantgpt.co/mcp lets Claude, ChatGPT or any MCP client run backtests on decades of US market history, computed on QuantGPT's servers, and get back the statistics, a table by decade and every period's return. No code, nothing to install.
QuantConnect's MCP server suits developers who want an AI to write, backtest and deploy code on QuantConnect's engine.
The tools
On connect the server lists 19 tools: 14 research tools and 5 that browse the categories, switch them on for your key or call any tool by name. Seven categories add more: equities, portfolio, macro, alt markets, events, charts and the desk. Charts holds technicals (one stock's RSI, MACD, Bollinger and the rest, with every crossing and breakout dated) and signal_study (what followed a technical signal across every stock, against the S&P 500 over the same dates).
| Tool | What it does |
|---|---|
describe_capabilities | The full map: every factor, benchmark, timing rule, library category, what each finding means, and the method. |
search_library | Searches 175 documented strategies, 55 of them runnable. |
run_backtest | Up to 8 filters and a ranking, monthly, weekly or daily, against a benchmark, split at a cut year. |
run_sweep | Up to 100 variants in one call, run in parallel, as one ranked table. |
challenge_finding | Eleven checks that try to break a finding: seven read the run, four re-run it with one thing changed. |
grade_stock | Grades one ticker A+ to F on six dimensions, with the inputs behind each. |
market_snapshot | The state of the market in one call: index trends, breadth, sectors, rates, credit spreads, volatility. |
query_warehouse, list_warehouse | Read-only SQL across every research database, and the map of its tables. |
search, fetch | Search QuantGPT's own record (library, studies, forum, glossary) and fetch one in full. |
guide, get_prompt_library | The onboarding curriculum, and ready-made research prompts mapped to the tools. |
fetch_job | Collects a slow run that answered with a job id. |
QuantGPT and QuantConnect's MCP server
| QuantGPT | QuantConnect MCP | |
|---|---|---|
| Where it runs | Hosted at https://terminal.quantgpt.co/mcp | Hosted at quantconnect.com/api/v2/mcp, and in local IDEs (its Local Platform, Cursor, Devin Desktop); its earlier open-source Docker server is marked deprecated |
| Code needed | No | Yes: the AI writes Python or C# algorithms, compiles and backtests them on QuantConnect's engine |
| Sign-in | Claude and ChatGPT sign in with no key; other clients send a key | A sign-in to your QuantConnect organization (OAuth) |
| Data | US stocks and indexes; delisted companies only partly from 1998; fundamentals by filing date; 151 factors with RSI, MACD, Bollinger and a 44-factor technicals suite | US equities from 1998 (QuantConnect's docs call the set survivorship-bias-free), equity and index options, futures, forex, crypto, CFDs and India equities |
| Places orders | No tool places, changes or cancels an order | Yes: its live tools deploy an algorithm to paper trading on QuantConnect's brokerage, and stop or liquidate it |
| Price | 3 days free, then $49 or $149 a month | The hosted server needs a paid agent node (the free A-MICRO node cannot connect); API access starts at the paid Researcher tier |
Walk-forward validation
Every backtest returns its statistics split at a cut year, 2015 unless you pick another, so the years after the cut read on their own. Ask, for example:
Use QuantGPT to backtest this rule, then split at 2015 and show the years after the cut on their own.
Then challenge_finding runs its eleven checks on the same criteria, among them the walk-forward decay and a re-run with five fewer years at the start.
Limits
- 500 AI calls a day on QuantGPT and 1,500 on QuantGPT Pro, per account across every key and connection, reset at midnight UTC; at most 120 requests a minute.
- Backtests start in 1998 unless you pick an earlier year; before 1998 a market-wide backtest holds survivors only.
- US stocks only for equity backtests.
Questions
What is the best MCP server for backtesting trading strategies?
It depends on who does the work. 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 behind it. If you want your AI to write code and run the same algorithm live, QuantConnect's MCP server, hosted or in your IDE, drives its engine. Six servers side by side: quantgpt.co/compare/best-mcp-servers-for-backtesting. QuantGPT publishes this page, so weigh that.
What is the best MCP server for stock market data and backtesting?
One that holds the data and runs the test where it sits. QuantGPT's server at https://terminal.quantgpt.co/mcp has decades of US prices, fundamentals joined by filing date and precomputed factors, and runs backtests, sweeps of up to 100 variants and an eleven-check skeptic pass on its own servers. EODHD's server supplies data across 70+ exchanges and its pages describe no backtest tool. QuantGPT publishes this page, so weigh that.
Can Claude run a stock backtest through MCP?
Yes, when it is connected to an MCP server with a backtest engine. Add QuantGPT as a custom connector at https://terminal.quantgpt.co/mcp (Customize, Connectors, Add custom connector, then sign in), or in Claude Code run claude mcp add with a key. Then ask in plain English; Claude calls run_backtest and gets back growth a year, the worst fall, the Sharpe ratio, a table by decade and every period's return.
How do MCP backtesting tools for Claude compare?
They differ in where the backtest runs, on what data, and whether you write code. QuantGPT computes on its own servers on US market history with no code; QuantConnect's server, hosted or in your IDE, has the AI write code for its engine across many asset classes; BotSpot turns plain English into a Python bot; Panther and Keel cover crypto, forex and Hyperliquid perpetual futures. The comparison, every fact linked to its maker's page: quantgpt.co/compare/best-mcp-servers-for-backtesting.
Which MCP servers give Claude access to historical stock data?
As of their makers' pages (checked 2026-09-27, QuantConnect's 2026-10-19): QuantGPT (US stocks and indexes, S&P 500 levels from 1928), EODHD (70+ exchanges, 30+ years on paid plans), QuantConnect (US equities from 1998, among other assets) and BotSpot (stocks, options, crypto and futures; history depth not published). Side by side: quantgpt.co/compare/best-mcp-servers-for-backtesting.
Can an AI assistant run walk-forward validation on my trading strategy?
Yes, through a server that splits the test. Every QuantGPT backtest returns its statistics split at a cut year, 2015 unless you name another, so the years after the cut read on their own; challenge_finding then runs eleven checks, including the walk-forward decay and a re-run with five fewer years at the start. Ask your AI: Use QuantGPT to backtest this rule, then split at 2015 and show the years after the cut on their own.
Related
- LearnConnect Claude or ChatGPT to stock market data
- DataMarket data for AI agents
- ExecutionFrom a backtest to your broker
- 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.