lunchmoney-readonly-mcp
A read-only MCP server for Lunch Money that provides secure, retrieval-only access to financial data through ChatGPT Business, with OAuth authentication and enforced read-only restrictions.
README
Read-only Lunch Money MCP for ChatGPT
A single-user remote MCP server for Lunch Money, designed for ChatGPT Business and deployed on Cloudflare Workers. It uses Google OAuth only to identify an allowlisted user; the Lunch Money API token is stored as a Cloudflare encrypted secret and is never returned to ChatGPT or an MCP tool.
This project was originally based on the bm1549/lunchmoney-mcp-cloudflare project and has been adapted into a read-only Lunch Money MCP template for Cloudflare Workers and ChatGPT. The bundled upstream implementation is retained under vendor-lunchmoney-mcp, including its original MIT license and attribution. Thanks to Brian Marks for the original work.
This repository is a template. It contains no live deployment URL, API token, Google OAuth secret, email address, or Cloudflare resource ID.
What makes it read-only
The server enforces read-only access in two independent layers:
createReadonlyServer()exposes a positive allowlist of 25 retrieval tools. Create, update, delete, upload, sync, refresh, split, grouping, and budget mutation tools are not registered.runWithReadonlyConfig()rejects any non-GETrequest to the Lunch Money API before a network request is made.
All exposed tools declare readOnlyHint: true, destructiveHint: false, and
openWorldHint: false.
Requirements
- A Lunch Money API token
- A Cloudflare account with Workers and KV available
- A Google Cloud OAuth 2.0 web client
- A ChatGPT Business workspace where you are an admin or owner
Deploy your own copy
-
Clone this repository, then install and verify dependencies:
npm install npm run typecheck npm run lint npm run test:readonly -
In
wrangler.jsonc, choose a unique Workernameand create the OAuth KV namespace:npx wrangler kv namespace create OAUTH_KVCopy the returned namespace ID into
REPLACE_WITH_OAUTH_KV_NAMESPACE_IDinwrangler.jsonc. -
Deploy once to obtain your Worker URL:
npx wrangler deploy -
Create a Google OAuth 2.0 Web application. Add this redirect URI, replacing the host with your Worker URL:
https://YOUR-WORKER.YOUR-SUBDOMAIN.workers.dev/authorize/callback -
Set the production secrets interactively. Never put their values in Git:
npx wrangler secret put GOOGLE_CLIENT_ID npx wrangler secret put GOOGLE_CLIENT_SECRET npx wrangler secret put ALLOWED_EMAILS npx wrangler secret put STATE_SECRET npx wrangler secret put LUNCHMONEY_API_TOKENALLOWED_EMAILSis a comma-separated list of Google accounts allowed to connect. The server fails closed when this is empty. -
Deploy again. Your MCP endpoint is:
https://YOUR-WORKER.YOUR-SUBDOMAIN.workers.dev/mcp
Connect ChatGPT Business
- In ChatGPT, open Workspace settings → Apps → Create.
- Enter the
/mcpendpoint above and choose OAuth authentication. - Review the custom-app warning, create the draft, then publish and enable it.
- Each user connects from Settings → Plugins → [your app] → Connect and signs in with an allowlisted Google account.
- Start a fresh Work chat, select the app, and verify the tool list before asking it to inspect Lunch Money data.
Before making the repository public
Run:
npm run check:public-release
Also review git status, confirm .dev.vars is untracked, and ensure the
Worker name, endpoint, Cloudflare resource IDs, and example documentation do
not identify your personal deployment.
If you are sanitizing an existing deployment, keep its configuration outside
the repository or in a gitignored wrangler.personal.jsonc file. Deploy that
private configuration with npx wrangler deploy --config wrangler.personal.jsonc.
Verification
The focused security test verifies the exact allowlist, its annotations, the absence of representative mutation tools, and rejection of outbound POST requests:
npm run test:readonly
See THIRD_PARTY_NOTICES.md for attribution details.
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。
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。
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。