plinth-fx-mcp
MCP server that provides real exchange-rate data from the European Central Bank, including latest rates, currency conversion, historical rates, and time series.
README
Plinth FX MCP
A small, production-clean Model Context Protocol server that gives any MCP client (Claude Desktop, Claude Code, Cursor, ...) real exchange-rate data — latest rates, currency conversion, historical rates, and time series — from the free, no-API-key Frankfurter service (European Central Bank reference rates).
It's free and MIT-licensed. The point isn't just "currency rates in your LLM" — it's a reference for what an MCP server should look like when you actually have to run it: rate-limited, retried, observable, and validated.
Why this one
Most example MCP servers are toys that fall over the first time an LLM calls a tool in a loop or the upstream hiccups. This one ships the boring stuff that keeps a server alive in production:
- Outbound rate limiting — a token bucket caps how fast we hit the upstream, so an over-eager agent can't get you throttled.
- Exponential backoff with full jitter — transient
429/5xx/network errors are retried with capped, jittered delays that honorRetry-After. - Structured JSON logging to stderr — one JSON object per line, ready for a log shipper. Never writes to stdout (that's the JSON-RPC channel — polluting it corrupts the session).
- Graceful shutdown — SIGTERM/SIGINT close the HTTP client cleanly.
- Input validation on every tool argument — bad currency codes, dates, or amounts fail fast with a clear, model-recoverable error instead of a stack trace.
- A real health check tool — reports version and live upstream reachability.
- Optional API-key gate — off by default (the upstream is public); flip on with one env var if you expose this server to others.
Quickstart (one command)
With uv installed, no clone or install needed:
uvx plinth-fx-mcp
Or with pip:
pip install plinth-fx-mcp
plinth-fx-mcp
The server speaks MCP over stdio. It's meant to be launched by an MCP client, not run interactively — but the command above will start and block, logging a startup line to stderr, which confirms it's healthy.
Use with Claude Desktop / Claude Code
Add this to your MCP client config (for Claude Desktop:
claude_desktop_config.json; for Claude Code: claude mcp add or your
project's .mcp.json):
{
"mcpServers": {
"plinth-fx": {
"command": "uvx",
"args": ["plinth-fx-mcp"]
}
}
}
Then ask: "Convert 250 USD to EUR, and show me how EUR/USD moved over the first week of January 2024."
Tools
| Tool | What it does |
|---|---|
latest_rates(base="EUR", symbols=None) |
Latest rates for a base currency, optionally filtered. |
convert(amount, from_currency, to_currency) |
Convert an amount at the latest rate (computed locally). |
historical_rate(date, base="EUR", symbols=None) |
Rates for a base currency on a past date. |
time_series(start, end, base="EUR", symbols=None) |
Rates over a date range, one entry per published day. |
list_currencies() |
Every supported currency code with its display name. |
health_check() |
Server version + live upstream reachability. |
Every data tool returns a JSON object with ok: true and the payload, or
ok: false with error and message on failure.
Configuration
All optional. Defaults are tuned for the free public upstream.
| Env var | Default | Purpose |
|---|---|---|
PLINTH_API_KEY |
(unset) | If set, every data tool requires a matching api_key arg. |
PLINTH_LOG_LEVEL |
INFO |
DEBUG / INFO / WARNING / ERROR. |
PLINTH_RATE_LIMIT_REQUESTS |
5 |
Max upstream requests per window. |
PLINTH_RATE_LIMIT_WINDOW_SECONDS |
1.0 |
Rate-limit window length. |
PLINTH_MAX_RETRIES |
4 |
Retries on transient upstream failures. |
PLINTH_BACKOFF_BASE_SECONDS |
0.5 |
Base for exponential backoff. |
PLINTH_BACKOFF_MAX_SECONDS |
20.0 |
Backoff cap. |
PLINTH_REQUEST_TIMEOUT_SECONDS |
15.0 |
Per-request timeout. |
Development
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
Tests mock the network (via respx), so they pass fully offline.
Want the whole toolkit?
This repo is one polished example. If you're building MCP servers for real and want to skip the yak-shaving:
→ MCP Server Starter Kit
Python and TypeScript templates, more worked examples, Dockerfiles, and a deploy + directory-publishing guide (how to actually get your server listed in the MCP registries). Same SRE hygiene as this repo, batteries included.
More from the brand at plinthkit.vercel.app.
License
MIT © Plinth. Exchange-rate data from Frankfurter, sourced from European Central Bank reference rates.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。