I Asked Claude Where to Invest $10,000. Here’s What Happened

A friend of mine has had $10,000 sitting in a savings account for eight months.

She’s not indecisive by nature. She runs financial models for a living.

But every time she sits down to actually invest it, the same thing happens. Fifteen browser tabs. Two YouTube videos titled “best stocks for 2026.” A note on her phone called “research” that never gets past the second bullet point.

Last month she asked what I’d do with it. I didn’t give her an answer.

I opened Claude instead.

Not to ask “what should I buy.” That question gets you a hedge and a shrug from any honest source, human or AI. I asked something more specific: connect to real market data, pull actual fundamentals and technical readings across a handful of assets, and build a $10,000 allocation you can defend with numbers instead of a feeling.

Here’s what came back, and why it’s a different exercise than what most people mean when they say “I asked AI where to invest.”

If you’re:

• sitting on cash you keep meaning to invest and never quite do,
• building a tool that needs an AI to reason over live financial data,
• or just curious what “AI-assisted investing” actually looks like once you remove the hallucinated numbers,

this is worth reading.

The problem with asking an AI for stock advice

My friend had actually tried this already, weeks before she asked me. She’d typed “what should I invest $10,000 in” into a chatbot and gotten three paragraphs back that sounded confident and said nothing she could act on.

That’s the default outcome. Ask ChatGPT or Claude the same question with no data connection, and you’ll get an answer that sounds reasonable and means nothing.

The model doesn’t know today’s price. It doesn’t know this week’s earnings. It doesn’t know whether the stock it just recommended is up 40% since its training cutoff or down 40%.

It’s pattern-matching against a snapshot of the internet that’s already stale.

Developers building financial tools run into this constantly:

• The model recommends a P/E ratio that was accurate 18 months ago
• It describes a company’s growth story without knowing the last two earnings reports
• It gives generic “consider diversifying” advice because it has no portfolio-specific numbers to reason over

None of that is a reasoning problem. It’s a data problem.

An AI is only as good as the numbers you hand it

Claude can reason well about markets. Sector rotation, valuation trade-offs, risk allocation, it understands the concepts fine.

What it can’t do on its own is know what MSFT trades at today, or what Johnson & Johnson’s current dividend payout ratio is, or how far gold has pulled back from its 2026 highs.

Connect it to a live data feed, and the reasoning suddenly has something real to work with.

Grounding the analysis in real data

For this, I used EODHD’s API, which exposes fundamentals, historical prices, and technical indicators for over 150,000 tickers through a set of REST endpoints (and, through MCP, directly callable by Claude inside a conversation).

Instead of asking Claude to “recommend stocks,” I asked it to pull current fundamentals and technical readings for a shortlist of assets across different risk buckets, then build a $10,000 allocation from what it found.

Not a hot take. A structured pull of P/E ratios, dividend yields, revenue growth, 52-week ranges, and price trend, followed by an allocation decision that had to reference those specific numbers.

Connecting Claude to EODHD’s MCP server, step by step

Before pulling any numbers, Claude needs a live connection to EODHD. Here’s the setup:

1. Get a free API key. Register at eodhd.com and grab your key from the dashboard. The free plan covers everything in this article.
2. Pick the right endpoint version. EODHD ships two: v2 uses OAuth 2.1 and is the one to use in Claude Desktop or claude.ai, since you never paste a key anywhere. v1 uses an API key in the URL and fits Claude Code, ChatGPT, or a custom agent.
3. Add the connector. In Claude, open Settings, go to Connectors, and add a custom connector pointing at https://mcp.eodhd.com/v2/mcp.
4. Authenticate. Approve the OAuth prompt from EODHD. No key copying, no environment variables.
5. Enable it in a chat and ask for something specific. “Pull MSFT’s current P/E ratio and 14-day RSI through EODHD” is a better first test than “tell me about Microsoft.” The first forces a tool call. The second lets the model fall back on memory.

That connector exposes 75 read-only tools: fundamentals, historical and intraday prices, technical indicators, options, macro data, news, and ESG scores, all sharing one schema instead of five.

Pulling the numbers

Here’s a simplified version of what a plain REST call looks like behind the scenes, in case you’re wiring this into your own script instead of a chat connector. Fundamentals and technical indicators, per ticker, through EODHD’s REST API:

import requests
API_TOKEN = "YOUR_API_KEY"
TICKERS = ["VOO.US", "MSFT.US", "JNJ.US", "SHY.US"]
def get_fundamentals(ticker):
    url = f"https://eodhd.com/api/fundamentals/{ticker}"
    params = {"api_token": API_TOKEN, "fmt": "json"}
    return requests.get(url, params=params).json()
def get_rsi(ticker, period=14):
    url = f"https://eodhd.com/api/technical/{ticker}"
    params = {
        "api_token": API_TOKEN,
        "function": "rsi",
        "period": period,
        "fmt": "json",
    }
    return requests.get(url, params=params).json()
for ticker in TICKERS:
    highlights = get_fundamentals(ticker).get("Highlights", {})
    rsi = get_rsi(ticker)
    print(ticker, "P/E:", highlights.get("PERatio"), "| latest RSI:", rsi[-1]["rsi"])

From here you can build:

• a screener that flags over- or under-valued positions in a watchlist
• a rebalancing agent that checks allocation drift monthly
• a research assistant that only speaks in numbers it just fetched, not numbers it remembers

That’s the actual unlock. Not “AI picks stocks,” but “AI reasons over verified, current data and shows its work.”

Why MCP instead of just calling the API

The script above works fine on its own. So why route through MCP at all?

Because with a plain REST call, you own the integration. You write the request, handle the response shape for each of EODHD’s endpoint categories, and update your code whenever a field changes.

With MCP, that work already lives on EODHD’s server. Claude sees a named tool with a defined schema, not a URL to remember. Ask for a stock’s fundamentals and technical indicators in the same sentence, and the model can call both tools and combine the results without you writing the orchestration logic.

Pros

• No integration code to maintain on your side
• The model can chain several tool calls (fundamentals, technicals, news sentiment) in one turn
• 100+ pages of EODHD’s own documentation are exposed as MCP resources, so the model can look up a field definition before using it

Cons

• Less control over exact request parameters than a hand-written call
• Adds a dependency on EODHD’s MCP server staying up, on top of the API itself

Best for: conversational research and prototyping. If you’re building a backend service with strict latency or rate-limit requirements, calling the REST API directly still makes sense there. MCP is for the layer where Claude is doing the reasoning.

Why EODHD instead of another data provider

Fundamentals from one vendor, technicals from another, news sentiment from a third: that’s the usual state of a DIY financial data stack, and it’s most of the integration work in a project like this.

EODHD’s pitch is one account covering fundamentals, historical and intraday prices, technical indicators, options, macro data, and news, across more than 150,000 tickers and 70+ exchanges, with 30+ years of history. For this article, that meant one API key instead of three.

The MCP server matters here too. Most alternative data providers don’t expose one yet, so even with a cheaper API, you’d still be writing your own tool wrapper for Claude to use it.

Cons worth knowing: the free tier has rate limits, and true real-time US tick data sits behind a paid plan. If your use case is high-frequency trading infrastructure, this isn’t that. If it’s AI-assisted research, screening, or the kind of allocation work in this article, the free tier is enough to start.

The $10,000 allocation

Here’s what came out of it, built for a moderate risk tolerance and a multi-year horizon, not a trade.

1. Vanguard S&P 500 ETF (VOO): $3,500 (35%)

VOO closed the week near $702, inside a 52-week range of $578 to $716, and sitting close to its all-time high. Expense ratio: 0.03%. Trailing P/E around 27 to 28. Dividend yield near 1.1%.

The technical picture is a stock still in an uptrend but no longer cheap by any historical measure. The fundamental picture is unremarkable in the way a core holding should be: broad exposure to roughly 500 large US companies, no single-stock risk, and an expense ratio that barely registers.

Best for: the base of the portfolio, not the return driver.

2. Microsoft (MSFT): $2,000 (20%)

MSFT traded around $491, off its 52-week high of $554 but well above its 52-week low of $349. P/E near 28. Dividend yield under 1%, small but growing.

The fundamentals carry the case here: revenue growth projected near 18% this year and over 19% next year, driven by Azure and AI infrastructure demand. Analyst consensus sits at “Strong Buy” with price targets clustered well above the current price.

Cons: a premium valuation that already prices in a lot of that growth, and full exposure to whatever happens to AI capex spending if the cycle cools.

Best for: growth exposure without single-product concentration risk.

3. Johnson & Johnson (JNJ): $2,000 (20%)

JNJ traded near $260, close to its 52-week high, with a P/E around 30, a dividend yield around 2%, and a payout ratio near 60%.

That last number matters more than it looks. JNJ has raised its dividend for 64 consecutive years. A 60% payout ratio means there’s room to keep doing that even if earnings growth slows.

Best for: the defensive counterweight to MSFT’s growth bet. Healthcare demand doesn’t move with the Nasdaq.

4. Gold, via a physically-backed ETF: $1,500 (15%)

This is the position the fundamentals can’t fully explain, because gold doesn’t have earnings or a P/E ratio. It has a price, and the price tells the story.

Spot gold spent 2026 on one of its sharpest runs in decades, pushing past $5,000 an ounce for the first time before shedding roughly $1,200 in two trading days, its worst two-day drop since 1983. It has since stabilized in the low $4,000s.

Central bank buying and rate-cut expectations are still supportive. But a chart that goes vertical and then falls 20% in 48 hours is not a chart you build a full position around.

Best for: a hedge sized small enough that a further correction doesn’t wreck the portfolio, not a bet on $6,000 gold.

5. iShares 1–3 Year Treasury Bond ETF (SHY): $1,000 (10%)

SHY trades around $82 with a 30-day SEC yield near 4%, an effective duration under two years, and volatility close to zero relative to equities.

There’s no technical or fundamental debate to have here. It’s the position that lets you rebalance into a drawdown instead of selling into one.

Best for: dry powder that earns a real yield while it waits.

If you’re building anything similar, the API tier that made this fast enough to actually run was EODHD’s, mostly because fundamentals, historical prices, and technical indicators live behind one consistent interface instead of three different providers.

→ Get started with EODHD’s API

Looking ahead: what would change this allocation

None of this is a prediction. It’s the short list of numbers worth re-checking before the next rebalance, because each of these positions has a specific thing that would break its thesis.

Rate cuts are the biggest swing factor for the whole mix. Cuts tend to support equity valuations further and are part of why gold has stayed bid despite its pullback. They’d also lower SHY’s yield on reinvestment, which is the trade-off for holding something that safe.

MSFT is the position with the least room for error. Its case rests on Azure and AI infrastructure growth holding near the 18 to 19% analysts are currently projecting. A slowdown in that specific number matters more here than in any other position, because the valuation already assumes it happens.

JNJ’s dividend math is the thing to watch, not the price. A 64-year increase streak and a 60% payout ratio give real room. A patent cliff or a major litigation headline would show up in the fundamentals before it shows up in the dividend.

Gold is the position to reassess first if it moves again. After a run past $5,000 and a 20% two-day drop, another leg down toward the low $4,000s wouldn’t break the thesis at this sizing. A break below the support levels most technical desks are already watching would be the signal to look again, not to add more.

What this allocation is actually betting on

Two large, quality-tilted equity positions carrying most of the growth. One dividend aristocrat absorbing the volatility. A gold position sized like a hedge, not a conviction call. Cash-like treasuries doing nothing exciting on purpose.

None of it is a hot tip. All of it is a number Claude could point to and defend.

Tips for running this yourself

• Re-pull fundamentals monthly, not daily. P/E ratios and payout ratios don’t move fast enough to justify checking them every day, and checking daily just invites trading on noise.
• Ask Claude to show the actual numbers it pulled, not just its conclusion. If it can’t cite the specific P/E or RSI value behind a claim, it’s reasoning without grounding again.
• Set a rebalancing rule before you start, something like “flag any position more than 5 percentage points off target,” and let the agent check that one number instead of re-litigating the whole allocation from scratch each time.
• Keep the connector scoped to read-only tools. Every tool in EODHD’s MCP server is read-only, so there’s no path for the model to place a trade. That’s the right default here.
• Start on the free tier. Upgrade once you know which specific endpoints you’re actually using, not before.

Key takeaways

• An AI without live data access is guessing with good grammar. Connect it to a real feed and the reasoning becomes something you can actually check.
• Fundamentals and technicals tell different stories on the same asset. VOO and JNJ both look “expensive” by P/E, but for very different reasons and with very different downside profiles.
• The most useful output wasn’t a stock pick. It was a portfolio structure where every position had a stated job.

This is not financial advice, and I’m not a financial advisor. It’s a demonstration of what happens when you stop asking an AI to guess and start asking it to look something up first.

Need technical content for your fintech, API, or developer product?
I write data-backed tutorials like this one, grounded in real API calls instead of marketing copy.
→ See my work

Looking for technical content for your company? I can help — LinkedIn · kevinmenesesgonzalez@gmail.com