DEV Community

vdalhambra
vdalhambra

Posted on

How to give Claude real-time data access with MCP (no API keys)

#mcp #claude #tutorial #ai

The problem with Claude out of the box is obvious the moment you ask it anything time-sensitive: it doesn't know what happened after its training cutoff. Ask it for Apple's current stock price, a live RSI signal, or whether your site's security headers are misconfigured — and you get either a refusal or a hallucinated answer.

Model Context Protocol (MCP) fixes this. It's a standard that lets you connect external tools directly into Claude's context at inference time. No custom APIs, no middleware, no prompt engineering hacks. Just a server that Claude can call.

This tutorial walks through what MCP actually is, why it matters, and how to get live financial data into Claude in under five minutes using FinanceKit MCP.

What is MCP, exactly?

MCP (Model Context Protocol) is an open standard — think of it like a USB-C port for AI models. It defines a consistent way for LLMs to call external tools, access resources, and retrieve context from the real world.

Under the hood, an MCP server is just a process that:

  1. Exposes a list of tools (functions with typed input/output schemas)
  2. Listens on a transport (stdio, SSE, or HTTP)
  3. Handles tool-call requests and returns structured responses

Claude acts as the MCP client. When you describe what you want, Claude figures out which tool to call, calls it with the right parameters, and incorporates the result into its response.

No polling. No RAG pipeline. No vector store. The data flows in real time, on demand.

Why this matters more than RAG

RAG (retrieval-augmented generation) is great for static document search. But financial data, site audits, and live APIs are fundamentally different: the data changes every second, and you can't pre-index it.

MCP solves the category of problems where:

  • The data is live (stock quotes, crypto prices, site uptime)
  • The computation is deterministic (RSI calculations, VaR modeling)
  • The output structure matters (you want a verdict, not a paragraph)

Install FinanceKit MCP (one command) 

claude mcp add financekit -- uvx --from financekit-mcp financekit
Enter fullscreen mode Exit fullscreen mode

That's it. uvx handles the Python environment. You don't need to manage dependencies, write config files, or set any API keys. FinanceKit pulls data from Yahoo Finance and CoinGecko — both free, no auth required.

To verify it's registered:

claude mcp list
Enter fullscreen mode Exit fullscreen mode

You should see financekit in the output.

Three prompts that actually work now

Once installed, open Claude and try these:

Prompt 1: Real-time stock quote

"What's the current price and 52-week range for NVDA?"

Claude calls get_stock_quote under the hood. You get the live bid/ask, day range, volume, market cap — sourced from Yahoo Finance at the moment you asked.

Prompt 2: Technical analysis with a verdict

"Run a full technical analysis on AAPL and tell me if the RSI and MACD are aligned for a long entry."

This triggers get_technical_analysis, which returns structured JSON with:

  • RSI (14-period) with an oversold / neutral / overbought verdict
  • MACD signal crossover status
  • Bollinger Band width and position
  • ADX trend strength

Claude then interprets all four signals together and gives you a coherent answer — not just raw numbers.

Prompt 3: Portfolio risk metrics

"Given a portfolio of 60% SPY, 30% QQQ, 10% BTC-USD, calculate the 30-day VaR at 95% confidence and the Sharpe ratio."

This calls calculate_portfolio_risk. The server pulls historical price data, runs the math, and returns VaR, Sharpe, Sortino, Beta, and a full correlation matrix. Claude explains what the numbers mean in plain English.

How it works under the hood

When Claude processes your prompt and decides it needs real-time data, here's what happens:

You → Claude (client)
         ↓  tool_call: get_stock_quote({ticker: "NVDA"})
    FinanceKit MCP Server
         ↓  fetches from Yahoo Finance
         ↑  returns {price: 875.20, change: +2.1%, ...}
Claude → incorporates result → responds to you
Enter fullscreen mode Exit fullscreen mode

The MCP server runs as a local process (stdio transport by default). Claude communicates with it via JSON-RPC messages. The server is stateless — each tool call is independent.

FinanceKit has 17 tools total:

  • get_stock_quote — real-time price, volume, market cap
  • get_technical_analysis — RSI, MACD, Bollinger, ADX with verdicts
  • get_crypto_price — CoinGecko live prices
  • calculate_portfolio_risk — VaR, Sharpe, Sortino, Beta, correlation
  • get_options_chain — live options data
  • get_earnings_calendar — upcoming earnings dates
  • ...and 11 more

All MIT licensed. No API keys. The full list is in the GitHub repo.

Add SiteAudit while you're at it

If you do any web development, add the companion server too:

claude mcp add siteaudit -- uvx --from siteaudit-mcp siteaudit
Enter fullscreen mode Exit fullscreen mode

SiteAudit MCP gives Claude 11 tools for website analysis: SEO audits, security header checks, Lighthouse scores, WCAG accessibility, competitor comparison, broken link detection, and robots.txt parsing.

Try: "Audit the security headers on example.com and tell me what's missing."

No install? Use the free tier

If you want to try the tools without installing anything locally, both servers are hosted on mcpize.com:

What's next

Once you understand MCP, the mental model shift is significant: Claude stops being a static knowledge base and starts being an orchestration layer that can call any tool you expose. The standard is still young, but the pattern is solid. Build a server, register it in one line, and your AI assistant suddenly knows things it couldn't before.

Both servers are open source and actively maintained. If you build something with them or find a bug, open an issue on GitHub.

Top comments (0)