If you've ever asked an LLM for today's USD-to-EUR rate, you already know the problem: these models serve up confident nonsense because they're working from training data that stopped updating months or years ago. For anything financial, that's a dealbreaker. Exchange Rate API just shipped two integrations designed to fix this specifically for AI agents running Claude Code, Cursor, DeepSeek, and any other MCP-compatible client.

The Core Problem

Large language models excel at understanding intent and formatting output, but they absolutely tank when asked for current numeric data. Ask GPT-4o or Claude 3.7 what $1,000 USD equals in EUR right now, and you'll get a number that sounded plausible during training but is meaningless today. This isn't a bug—it's fundamental to how these models work. They need external tools for the deterministic calculation step. Exchange Rate API's solution: two open-source MIT-licensed packages that wire real-time exchange rates directly into agent toolchains. No more hallucinated conversions, no more "around 0.92" hand-waving when precision matters.

MCP Server for Claude Code and Cursor

The @exchangerateapi/mcp-server package implements the Model Context Protocol standard, giving any MCP-compatible AI client access to live currency data over stdio. Installation is straightforward—for Claude Code, a single command adds the server: claude mcp add exchangerateapi -- npx -y @exchangerateapi/mcp-server. Cursor and Claude Desktop users drop equivalent JSON into their MCP config files (~/.cursor/mcp.json or the desktop app's configuration). Restart the client, and four new tools become available: get_exchange_rate, convert_currency, get_historical_rates, and list_currencies. Notably, the first two work without an API key; historical rates and authenticated endpoints require one.

DeepSeek Function-Calling Package

For Python developers working with DeepSeek's models, the exchange-rateapi-deepseek package provides tool schemas, an agent wrapper, and a CLI. One-shot queries run via exchange-rateapi-deepseek --ask "What is 2500 USD in EUR right now?", or developers can import DeepSeekCurrencyAgent for library usage. Under the hood, if deepseek-chat decides to call convert_currency, the package executes it against Exchange Rate API and passes JSON results back—DeepSeek then produces a final answer citing the actual rate used.

Real-World Use Cases

The integration targets three concrete scenarios that developers have been requesting: invoice review (matching USD invoices to EUR books), travel planning (accurate JPY-to-home-currency conversion for trip budgets), and cross-border quotes (generating BRL figures from USD SOWs for Brazilian clients). Each case requires the LLM to handle intent understanding while a deterministic tool handles the numeric step.

Under the Hood

Both packages expose identical operations regardless of interface—MCP, DeepSeek function calling, or direct REST API. Rates come from Reuters (Refinitiv) and interbank market feeds, covering 160+ currencies at mid-market rates. Four endpoints power everything: GET /api/rate for current rates (no key), convert_currency for amount calculations (client-side math on the rate, no key), GET /api/historical-rates (key required), and GET /api/v1/symbols to list supported currencies (no key).

ChatGPT Desktop Gets Love Too

ChatGPT Desktop added MCP support in late 2025, so the same server configuration works there. For web-based chat at chatgpt.com, Exchange Rate API offers an OpenAPI spec for Custom GPT Actions and maintains /llms.txt documentation that browsing-enabled LLMs can parse directly.

Key Takeaways

  • LLM hallucination on financial data is a solved problem with proper tool integration
  • MCP server works across Claude Code, Cursor, Claude Desktop, and ChatGPT Desktop
  • DeepSeek Python package enables one-shot queries or library-based agent development
  • Two of four tools work without API keys (get_exchange_rate, list_currencies)

The Bottom Line

These integrations represent exactly how AI agents should handle real-world data: use LLMs for reasoning and language, external tools for ground truth. If you're building anything that touches currency— invoicing systems, travel planners, international pricing calculators—you need this in your agent stack yesterday.