# Best AI trading bot setup (paper trading first)

Updated: 2026-10-24 | Cluster: Data and analysis | Canonical: https://theailoadout.com/best-ai-setup/ai-trading-bot/

> Educational only, not financial advice. Trading involves risk of loss, including losing more than you put in with leverage. Start with paper trading. We may earn affiliate commissions from some links, which never change our picks. Nothing here is a performance or profit claim.

The safest way to build an AI trading agent today is Claude Code connected to Alpaca's official MCP server on a free paper trading account. The MCP server defaults to paper trading, so the agent can check positions, pull market data and place practice orders for stocks, options and crypto without real money. Use TradingAgents as the research bot that proposes buy, hold or sell calls, and log those calls against the paper account for months before you trust anything. For crypto, see the crypto trading bot setup with Freqtrade. This page is educational only and is not financial advice.

| Slot | Pick |
|---|---|
| Tool | [Alpaca (paper trading account)](https://docs.alpaca.markets/us/docs/trading-api-mcp) |
| Model | [Claude Opus 5.5 (frontier) or Qwen 3.5 27B (local)](https://theailoadout.com/models/claude-opus-5-5/) |
| Skill | [Alpaca MCP server (paper mode)](https://docs.alpaca.markets/us/docs/trading-api-mcp) |
| Persona | [The Skeptical Trading Analyst](https://theailoadout.com/best-ai-setup/ai-trading-bot/#copy-this-setup) |
| Bot | [TradingAgents](https://theailoadout.com/bots/tradingagents/) |

## Why each pick
- Tool: Alpaca's docs say a free paper trading account is enough to start with its Trading API MCP server, which works with Claude Desktop, Claude Code, Cursor, VS Code and Gemini CLI. The server covers account and positions, orders (market, limit, stop, trailing stop, brackets, multi-leg spreads), market data, screeners and watchlists. It defaults to paper and only trades real money if you set ALPACA_PAPER_TRADE=false with live keys.
- Model: Most of the work here is writing and checking strategy and backtest code, so use a strong coding model. Opus 5.5 at max effort in Claude Code topped Terminal-Bench 4.0 per GenAI Brief in mid October 2026. To keep strategies and data on your own machine, run Qwen 3.5 27B or Gemma 4 26B in Ollama or LM Studio on a 24 to 32 GB computer (see our run AI locally setup).
- Skill, Alpaca MCP server (paper mode): Lets the agent read your paper account and place practice orders from chat. Keep keys in the MCP client env block, never in chat, as Alpaca's docs say.
- Skill, Backtest before every change: Ask the agent to write a backtest for every rule it suggests and to report the period, number of trades and worst drawdown. Freqtrade and Jesse ship backtesting if you would rather not write your own.
- Persona: LLMs sound confident about markets. This persona forces risk limits, backtests and plain language about uncertainty.
- Bot: TradingAgents is an open-source multi-agent LLM framework that outputs a buy, hold or sell decision per ticker and date. It does not place orders, per an InvestingStack audit, which is what you want while you are learning: it proposes, you check it against paper results.

## Comparison
| Option | Best for | Model(s) | Price | Free plan |
|---|---|---|---|---|
| Alpaca MCP + Claude Code | Stocks, options and crypto, paper first | Any MCP client | Paper account free; Algo Trader Plus data $99/mo | Yes |
| TradingAgents | Research decisions, no orders | Your LLM of choice | Free code; you pay LLM calls | Yes |
| Freqtrade | Crypto strategy bots with dry-run | FreqAI (ML) | Free (GPL-3.0) | Yes |
| Hummingbot | Crypto market making and arbitrage | Condor AI harness | Free (Apache 2.0) | Yes |

## Persona prompt
```
You help me research and test trading ideas on a PAPER account only. Never place a live order and never suggest disabling paper mode. For every idea: state the rule in one sentence, write a backtest, report the test period, number of trades, win rate and worst drawdown, and list three ways the result could be overfit. Every order you propose must include a position size of no more than 2 percent of the paper account and a stop. Say "I don't know" when you don't. Never promise or predict returns. Remind me this is not financial advice when I ask whether to buy or sell.
```

## What changed
- 2026-10-24: Page launched with initial picks. Prices and features were checked on the official pages listed under Sources.

## FAQ
### What is the best AI trading bot?
There is no bot with reliable proven returns, so the best setup is one you can test safely. We pick Claude Code with the Alpaca MCP server on a paper account, plus TradingAgents for research. Not financial advice.

### Can an AI trading bot make money?
We make no profit claims. Backtests and bot track records are hypothetical and often overfit, and real trading can lose money. Treat any bot promising fixed or guaranteed returns as a likely scam.

### Is there a free AI trading bot?
Yes. TradingAgents, Freqtrade and Hummingbot are free and open source, and Alpaca paper accounts are free. You pay for hosting and any LLM calls.

### What is paper trading?
Paper trading runs your strategy against live prices with fake money, so you can see how it behaves without risking any.

## Sources
- [Alpaca docs: Trading API MCP server](https://docs.alpaca.markets/us/docs/trading-api-mcp)
- [Alpaca market data plans](https://alpaca.markets/data)
- [TradingAgents repository (GitHub)](https://github.com/TauricResearch/TradingAgents)
- [InvestingStack: 20 AI trading agents, an open-source audit](https://investingstack.ai/research/open-source-ai-trading-agents-audit)
- [Freqtrade README (GitHub)](https://github.com/freqtrade/freqtrade)
- [Hummingbot.org](https://hummingbot.org/)
- [TradingAgentStack: open-source AI trading bots (Oct 23, 2026)](https://tradingagentstack.com/open-source-ai-trading-bots/)
- [GenAI Brief: best LLM for coding, October 2026](https://genaibrief.com/models/best-llm-for-coding/)