Lab MCP › MCP backtesting servers compared
How do MCP backtesting tools for Claude compare?
An MCP server (Model Context Protocol) lets Claude call real tools instead of improvising backtest code: market data, a simulation engine, statistical checks. The public servers do very different jobs. TradingView and QuantConnect run your strategy on their own servers, inside your account. Alpaca connects Claude to a broker. MaverickMCP, AI-Trader, backtester-mcp and the Strategy Arena Lab run on your machine, and they differ mostly in how they check a result on data the strategy has not seen.
Each line below was checked on 22 November 2026 against the publisher’s own page or repository, linked in the text. Details not stated there were left out. Features change: check the source before you choose.
TradingView’s official MCP server
TradingView publishes an official remote MCP server (https://mcp.tradingview.com/mcp, OAuth 2.1, 38 tools in 9 categories): watchlists, market data, screeners, news, documents, calendars, alerts and Pine scripts. With run_pine_backtest, Claude runs a Pine strategy in TradingView’s Strategy Tester and reads the report (net profit, profit factor, drawdown, Sharpe, Sortino, buy-and-hold return) and up to 200 recent trades. Claude can also validate, save and edit your Pine scripts. If you already live on TradingView, it is the most direct way to bring your charts into a conversation with Claude.
- Paid plan and call quotas. Every tool call counts against your plan: 10 calls a day on Essential (or a legacy plan), 20 on Plus, 50 on Premium, 500 on Ultimate. The optional AI subscription (AI Copilot + MCP) raises any paid plan to 500 calls a day and 50 a minute. The history depth and Deep Backtesting available also depend on the plan.
- Personal use. The terms reserve it for you, through AI clients you use yourself: no team or company use, no sharing or reselling the data, no bulk downloads, no AI training, and no automated trading or broker orders placed by an AI client.
Source: tradingview.com/mcp/docs (usage limits, tools reference, terms of use).
The comparison table
| MCP server | Where backtests run | Data | Check on unseen data | Places orders | Licence | Price |
|---|---|---|---|---|---|---|
| TradingView MCP (official) | TradingView’s servers (Strategy Tester) | TradingView market data, news, documents; depth by plan | Strategy Tester report on the chosen range | No (forbidden by the terms) | Proprietary, personal use | Paid TradingView plan; 10 to 500 calls a day |
| QuantConnect Remote MCP | QuantConnect cloud (LEAN) | QuantConnect datasets | Backtests and parameter optimization; your own split in code | Yes (live algorithm deployment) | Apache-2.0 (repository); cloud platform under its terms | Paid agent node required (the free A-MICRO node cannot connect) |
| Alpaca MCP Server | None (no backtesting engine) | Alpaca stocks, options, crypto: quotes, bars, news | Not its job | Yes (paper by default, live when set) | MIT | Free server; free paper account |
| MaverickMCP | Your machine (Python) | Yahoo Finance via yfinance, no API key for core tools | Rolling walk-forward, Monte Carlo on trades | No (local records only) | MIT | Free |
| AI-Trader | Your machine (Backtrader) | Fetchers for US and Taiwan stocks, crypto, forex; your CSV | Standard backtests | Not described | GPL-3.0 | Free |
| backtester-mcp | Your machine (NumPy, Numba) | Synthetic series, repository datasets, your CSV or Parquet | Bootstrap Sharpe interval, deflated Sharpe, PBO, walk-forward | No order tool among its 13 tools | Apache-2.0 | Free |
| atilaahmettaner/tradingview-mcp (independent) | Your machine when self-hosted, or hosted | Third-party screener data, Yahoo Finance | Walk-forward with a ROBUST to OVERFITTED label | No | MIT | Free self-hosted; hosted from $9 a month |
| Strategy Arena Lab | Your machine (CPU, plus CUDA GPU on Windows), 127.0.0.1 only | Local data vault | Final period sealed before any computation, every attempt counted, one opening: PBO, deflated Sharpe, Hansen’s SPA against buy-and-hold, CPCV, walk-forward, Monte Carlo; can answer INCONCLUSIVE | No | MIT connector; Lab app not open source | Free; optional server certificate: 1 credit |
Each server in a few lines
QuantConnect Remote MCP
A hosted endpoint (https://www.quantconnect.com/api/v2/mcp) that connects Claude to your QuantConnect organization: create and edit Python or C# projects, compile, run cloud backtests, optimize parameters and deploy live algorithms. It requires a paid agent node such as A1-1 or A16-32; the free A-MICRO node cannot connect. Sources: GitHub, documentation.
Alpaca MCP Server
Alpaca’s own MCP server for its Trading and Market Data APIs: quotes, historical bars, option Greeks, news, account data, and orders on stocks, options and crypto. It runs locally over stdio and uses paper trading unless you switch it to live. It has no backtesting engine: Claude can read historical bars and analyse them. Source: GitHub.
MaverickMCP
A local stock analysis server in Python (version 1.2.1: 38 core tools, 12 optional backtesting tools, 3 optional research tools). Data comes from Yahoo Finance through yfinance. The backtesting extra adds single-strategy backtests, grid optimization, rolling walk-forward, a Monte Carlo that resamples trades and portfolio backtests. It does not submit orders to a broker. Source: GitHub.
AI-Trader
A config-driven backtesting framework built on Backtrader, with more than 20 built-in strategies and data fetchers for US stocks, Taiwan stocks, crypto and forex. Its built-in MCP server lets Claude run backtests, fetch market data and list strategies. Source: GitHub.
backtester-mcp
A local engine installed with pip install backtester-mcp, built around one question: is this result overfit? Its 13 MCP tools include a full validation pipeline with a pass or caution verdict, bootstrap Sharpe confidence interval, deflated Sharpe ratio, PBO, walk-forward, three fill scenarios with estimated spread and impact, Bayesian optimization and a local DuckDB store of runs. It works on synthetic series or your own CSV or Parquet files. Source: GitHub.
atilaahmettaner/tradingview-mcp
An independent open-source project, not affiliated with TradingView Inc., listed in the official MCP Registry. Its 39 tools cover screeners, multi-timeframe technical analysis, Yahoo Finance prices, news, strategy backtests and walk_forward_backtest_strategy, a train and test split that returns ROBUST, MODERATE, WEAK or OVERFITTED. It needs no TradingView account and does not execute trades. Self-hosting is free; the hosted version starts at $9 a month. Source: GitHub.
The Strategy Arena Lab: the free, local complement
The Lab is a free desktop app for Windows and Linux with a local MCP server. Claude Code, Claude Desktop, Codex, Cursor, LM Studio or opencode connect over stdio or HTTP on 127.0.0.1 with a private token. Your AI writes rules in ArenaScript, backtests them on your PC and sweeps variants on the CPU or a CUDA GPU on Windows (1,000,000 variants in about 20 seconds on an RTX 4080 Super, Lab 235 measurement, legacy execution semantics); every kept row is replayed on the CPU reference engine. Your strategies, data and results stay on your machine.
- A judge your AI cannot steer. At the start of a session the final part of the window (30% by default over MCP) is sealed and the seal is written to disk before any computation. Tools that would read those bars are refused by name. Every valid script and GPU variant counts as an attempt, even in an abandoned session, and the total goes into the deflated Sharpe and the PBO of the single final proof.
- Measured before release. On pure noise, the judge said “survives” in none of 350 autoresearch sessions (11 October 2026).
- Back to TradingView. The Lab exports the exact Pine v5 of the judged strategy, and
compare_tradingview_tradeschecks your List of trades trade by trade: 234 of 234 trades matched in TradingView on the 6 leaderboard strategies (Lab 236 export, 8 October 2026). A free TradingView account is enough to paste the Pine. - No orders, no payments. No tool places an order, opens TradingView, publishes anything or starts a payment. The MCP connection, backtests, sweeps, Pine export and sealed proof are free; a server-signed certificate uses 1 credit, and paying never changes a verdict.
TradingView’s MCP and the Lab fit together: use TradingView’s MCP to explore your charts and scripts with Claude, and the Lab to search variants on your PC and see what holds on a period the search never touched, before you paste the Pine back into TradingView.
Strategy Arena also runs a public remote MCP server (https://strategyarena.io/mcp, no authentication, 45 tools) for public proofs and paper previews. Its server card states sealed: false: it opens no seal and gives no verdict; the verdict stays in the local Lab.
Which one fits what you want to do
- Talk to your TradingView charts and scripts: TradingView’s official MCP, on a paid plan.
- Develop and deploy algorithms in a cloud platform: QuantConnect Remote MCP.
- Give Claude access to a brokerage account: Alpaca MCP Server, in paper mode first.
- Run quick Python backtests on stocks locally: MaverickMCP or AI-Trader.
- Run statistical overfitting tests on your own data: backtester-mcp.
- Let an AI search strategies on your PC and judge the result on a sealed period, then export the Pine to TradingView: the Strategy Arena Lab, free.
No MCP at hand? Two free checks work in the browser: paste a Pine strategy into the free online judge (the last 20% of the candles stay sealed), or upload the List of trades exported from the Strategy Tester to get the free preview of the audit dossier.
Questions
Does TradingView have an official MCP server for Claude?
Yes. TradingView publishes an official remote MCP server at https://mcp.tradingview.com/mcp (OAuth 2.1, 38 tools in 9 categories), including Pine Script tools such as validate_pine_script and run_pine_backtest. It needs a paid TradingView plan, with a daily call quota from 10 (Essential) to 500 (Ultimate) that the AI subscription raises to 500 calls a day on any paid plan. Its terms limit it to personal use and rule out automated trading.
Which MCP backtesting servers run locally?
MaverickMCP, AI-Trader, backtester-mcp, atilaahmettaner/tradingview-mcp when self-hosted, and the Strategy Arena Lab run backtests on your own machine. TradingView’s official MCP and QuantConnect’s remote MCP run them on their own servers. Alpaca’s MCP server runs locally but is a broker connection with no backtesting engine of its own.
Which MCP servers check a strategy for overfitting?
Several offer a train and test split: MaverickMCP (rolling walk-forward and Monte Carlo) and atilaahmettaner/tradingview-mcp (walk-forward with a ROBUST to OVERFITTED label). backtester-mcp makes it its core: bootstrap Sharpe interval, deflated Sharpe ratio, PBO and walk-forward. The Strategy Arena Lab seals the final part of the history before any computation, counts every attempt and opens the sealed period once for the verdict. None of them can promise future results.
Can Claude place trades through these MCP servers?
Only two of them have order tools: Alpaca’s MCP server (paper trading by default, live trading when configured) and QuantConnect’s remote MCP, which can deploy live algorithms. MaverickMCP, backtester-mcp, atilaahmettaner/tradingview-mcp and the Strategy Arena Lab have none, and TradingView’s terms forbid letting an AI client place broker orders.
Can I use TradingView’s MCP and the Strategy Arena Lab together?
Yes, they do different jobs. TradingView’s MCP lets Claude read your charts, data and Pine scripts and run them in the Strategy Tester. The Lab is free and local: your AI searches variants on your PC, the judge counts the attempts and replays the chosen rule once on a sealed period, then the Lab exports the exact Pine to paste into TradingView, where you should find the same trades.
See also: connect an AI to the Lab over MCP · out-of-sample test on TradingView · audit dossier of a TradingView export
Research on past data. Every tool on this page evaluates rules on historical market data; a backtest does not predict future results. Strategy Arena gives no investment advice and does not trade for you. TradingView is a trademark of TradingView, Inc.; QuantConnect, Alpaca and the other names belong to their owners. Strategy Arena is not affiliated with them.