Personal Investing AI — AI Stack for Self-Directed Investors

What's in this pack

This is the stack for the self-directed investor — the person who runs their own brokerage account in stocks, ETFs, crypto, and the occasional options trade, and is tired of pasting earnings releases into a chat window one paragraph at a time. Not a Twitter list of 30 'AI alpha' bots. Every pick here does one specific job in the personal investing loop: pull market data, read filings with AI, monitor what you actually own, backtest before you trade, and keep enough of a paper trail that tax season is one afternoon.

The pack is built around one principle: AI accelerates the research; the trade and the tax record stay yours. That's why the data layer is OpenBB (open-source, you own the keys) instead of a paid terminal. That's why the portfolio layer is local-first (Wealthfolio, Ghostfolio) instead of a SaaS that sees every position you hold. And it's why the ledger is plain-text (Ledger) so you can hand a single file to your accountant in April.

The five layers

• Data sources — OpenBB, TWStockMCPServer — pull prices, fundamentals, filings, with API keys you control.
• Research / AI reading — Quant Analyst, TradingAgents, awesome-trading-agents — point agents at 10-Ks, earnings transcripts, and industry comps.
• Portfolio monitoring — Wealthfolio, Ghostfolio — track holdings, P&L, and allocation locally without uploading your book.
• Strategy & backtest — Backtrader, CCXT — write a strategy, replay it on historical data, then paper-trade on real exchanges before risking a dollar.
• Tax / ledger — Ledger — one plain-text file of every fill, ready for cost-basis math at year end.

Install in this order (data pipeline → filings AI → portfolio dashboard → strategy → tax ledger)

1. OpenBB — Open-Source Investment Research Platform — start here. OpenBB is a unified Python SDK and CLI that pulls market data, fundamentals, macro series, news, and filings from dozens of providers. Bring your own keys (free tiers exist for several). This is the data spine the rest of the pack reads from.
2. TWStockMCPServer — Taiwan Stock Data MCP Server — example of the MCP shape: an LLM-callable server that hands market data straight to your agent. If you trade non-US markets, this is the pattern to copy; if you don't, skim it once and move on. The point is that data layers are pluggable.
3. Claude Code Agent: Quant Analyst — once data is reachable, this is the agent you point at it. Valuation, DCF, scenario analysis, ratio screens, sensitivity tables. Feed it a fresh 10-K plus the last four quarters and ask 'what changed in the numbers'. Good at modeling, mediocre at judgment — the human still picks the assumption set.
4. TradingAgents — Multi-Agent LLM Financial Trading Framework — a research-grade multi-agent setup that simulates an analyst desk (fundamentals, sentiment, news, technicals, risk). Use it the way you'd use a junior team: it produces a memo, you decide. Do not wire it to a live order ticket on day one.
5. awesome-trading-agents — Trading Agents + MCP List — the index. Curated list of trading agent frameworks and MCP servers worth tracking — useful when picking 4 and 5's actual configuration, and for finding domain-specific MCPs (options chains, on-chain, alt-data) as new ones land.
6. Wealthfolio — Private Local-First Portfolio Tracker — your positions, on your machine, no cloud. Multi-account, multi-currency, manual entry or import. The right answer when you don't want a vendor to see your book.
7. Ghostfolio — Open Source Wealth Management & Portfolio Tracker — the self-hosted-server counterpart to Wealthfolio. Pick this if you want a web UI shared across devices (phone + laptop) and you're willing to run a Docker container. The two are alternatives — pick one based on the deploy story.
8. Backtrader — Python Algorithmic Trading Framework — the workhorse for strategy backtesting in Python. Write a strategy, replay on historical bars, get metrics (Sharpe, drawdown, win rate) before you risk capital. Pairs with OpenBB for the data feed.
9. CCXT — Universal Cryptocurrency Exchange Trading Library — if any part of your book is crypto, this is the integration layer. Unified API across 100+ exchanges for reading prices and placing orders. Start in read-only mode for monitoring; only enable trading keys after a long paper-trade window.
10. Ledger — Double-Entry Accounting via the Command Line — the tax/audit backbone. Every fill (buy, sell, dividend, fee, FX conversion, crypto trade) lands as a double-entry posting in a plain-text journal. At tax time you have one file your accountant can read in any decade, no proprietary export needed.

How they fit together

     Market data APIs
     (Yahoo, FMP, AV,
      exchange feeds)
            │
            ▼
      ┌─────────────┐         ┌──────────────────┐
      │   OpenBB    │ ←─MCP─→ │ TWStockMCPServer │
      │  (Python    │         │  (or other       │
      │   SDK/CLI)  │         │   regional MCP)  │
      └─────┬───────┘         └────────┬─────────┘
            │                          │
            └──────────┬───────────────┘
                       ▼
             ┌─────────────────────┐
             │  AI research layer  │
             │  Quant Analyst      │ ← 10-Ks, earnings, comps
             │  TradingAgents      │ ← multi-agent memos
             │  awesome-trading-   │ ← index of more MCPs/agents
             │   agents (catalog)  │
             └──────────┬──────────┘
                        │
           memos / models / signals
                        │
     ┌──────────────────┴──────────────────┐
     ▼                                     ▼
Portfolio dashboards               Strategy / execution
(Wealthfolio  ◇  Ghostfolio)       (Backtrader for backtest,
 positions, P&L, allocation         CCXT for crypto venues)
     │                                     │
     └──────────────┬──────────────────────┘
                    ▼
             every fill / dividend
                    ▼
             ┌─────────────┐
             │   Ledger    │ ← plain-text journal,
             │ (plain-text)│   cost basis, tax export
             └─────────────┘

The critical join is the read-only AI on top of your own data: the agents (Quant Analyst, TradingAgents) read the OpenBB-fetched data and your Ledger journal; they propose memos and signals; you place the trade, and the trade is logged back into Ledger. The agents never own the order ticket or the tax file. Read-only AI on auditable data is the safe shape for personal investing.

Tradeoffs you'll hit

• Real-time vs end-of-day data — real-time stock and options feeds cost money (broker API or paid vendor). End-of-day is free or near-free. For research and backtesting, end-of-day is more than enough. For active intraday trading, you'll pay. Don't pay before you've proven you actually trade intraday.
• Data cost vs coverage — OpenBB free providers cover US equities and crypto well, international equities and options chains less so. Plan to layer paid keys (FMP, Polygon, etc.) only for the slices you actually use. Subscribing to everything 'in case' is a $200/month tax on nothing.
• Model bias on financial text — an LLM summarizing an earnings call will mirror the management tone. Adverse signals (rising churn, deferred capex) get smoothed into 'continued investment in growth'. Mitigation: ask for the numbers that moved, not the narrative, and cross-check against the 10-K body.
• Backtest overfitting — Backtrader will happily show a 60% annualized return on a strategy that's curve-fit to one regime. Out-of-sample testing, walk-forward analysis, and paper-trading on live data for at least one quarter are not optional. A backtest is a hypothesis, not a license.

Common pitfalls

• AI-written strategies with look-ahead bias — the single most common failure mode for LLM-generated backtests. The model writes a strategy that, on close inspection, uses tomorrow's data to make today's decision (e.g. referencing close[t] when only close[t-1] was available at decision time). Always inspect the time index in the strategy logic; reject any 'too good' Sharpe without staring at the entry condition for ten minutes.
• 24/7 crypto monitoring cost — CCXT can poll exchanges as fast as their rate limits allow, and running that on a $10 VPS will rack up API errors, ban risks, and occasional missed fills. For personal accounts, 1-minute polling is plenty for almost everything below intraday scalping. Don't engineer for HFT you'll never do.
• Tax jurisdiction differences — Ledger is a generic accounting tool; it doesn't know your country's wash-sale rules, holding-period brackets, or crypto-as-property treatment. The pack gives you the data; your accountant (or local tax software) applies the rules. Don't ship a tax return out of an LLM. Talk to a human.
• FMP / data provider rate limits — free tiers cap requests per minute and per day. Hit the cap during a Sunday-night research session and your dashboard goes blank. Cache aggressively (a local Parquet of daily bars is one line of code), and treat the provider as the slow side of your pipeline.
• Treating the AI output as a recommendation — none of this pack constitutes financial advice. The agents produce memos, models, and backtests; you make the investment decision and you carry the loss. If a position would make you panic at -20%, don't take it because an LLM said the DCF was attractive.

Disclaimer

This pack is editorial guidance about an AI-assisted personal-investing workflow. It is not investment, tax, or legal advice. Tools mentioned have their own terms of service, data licenses, and regional restrictions — review them before use. Backtest results are not predictive of future performance. Past performance is not a guarantee of future returns. Talk to a licensed advisor before any material investment decision.