opendata-mcp
Enables interaction with the OpenData Platform API by dynamically exposing all endpoints as MCP tools with typed input schemas and HTTP handlers.
README
opendata-mcp
A Model Context Protocol server for the OpenData Platform API. Tools are generated dynamically from the bundled OpenAPI 3.1 spec — every endpoint is exposed as an MCP tool with a typed input schema and a transparent HTTP handler.
- GET endpoints: 76 — exposed without auth (the API serves them unauthenticated).
- Write endpoints (POST / PUT / PATCH / DELETE): 23 — require a Bearer token via
OPENDATA_API_KEY. - Built on
@modelcontextprotocol/sdk(TypeScript) with stdio transport.
Install
Uses pnpm (pinned via packageManager in package.json so corepack enable is enough to get the right version).
pnpm install
pnpm run build
Configure
Copy .env.example → .env (or set in your MCP client config) and fill in:
OPENDATA_BASE_URL=https://api.tryopendata.ai # required at call time
OPENDATA_API_KEY=od_live_xxx # required for writes
# OPENDATA_READ_ONLY=true # optional: only register GET tools
# OPENDATA_TIMEOUT_MS=30000 # optional: per-request timeout
# OPENDATA_MAX_ATTEMPTS=3 # optional: total HTTP attempts (retries are bounded; see below)
# OPENDATA_LOG_LEVEL=info # optional: off|error|info|debug
Secrets are never read from the repo. .env is gitignored; .env.example ships placeholders only.
Run
# Dev (tsx, no build step)
pnpm run dev
# Production (after build)
pnpm start
The server speaks MCP over stdio — wire it into any MCP-capable client. Example client config (Claude Desktop / similar):
{
"mcpServers": {
"opendata": {
"command": "node",
"args": ["/absolute/path/to/opendata-mcp/dist/index.js"],
"env": {
"OPENDATA_BASE_URL": "https://api.tryopendata.ai",
"OPENDATA_API_KEY": "od_live_xxx"
}
}
}
}
Test
pnpm test # unit tests (mocked fetch — offline-safe)
pnpm run typecheck # tsc --noEmit
Live tests against the real API run only when OPENDATA_BASE_URL is set:
OPENDATA_BASE_URL=https://api.tryopendata.ai pnpm test
CI runs the unit suite on every push/PR (badge above). The live smoke job runs on a daily schedule (and on pushes to main) only if the repo has OPENDATA_BASE_URL (and optionally OPENDATA_API_KEY) configured as Actions secrets — it's continue-on-error: true so a transient upstream outage doesn't show the repo as red.
How tool generation works
At startup the server:
- Loads
openapi.json(OpenAPI 3.1, bundled). - Walks every
(path, method)pair (skippingdeprecatedoperations and skipping non-GET whenOPENDATA_READ_ONLY=true). - Converts each operation's path + query parameters (and JSON request body for writes) into a Zod input schema via
openapi-to-zod.ts. - Registers one MCP tool per operation. The tool name is the OpenAPI
operationId(sanitized to[a-zA-Z0-9_-]{1,64}); the description isMETHOD path (auth?) — summary — description. - On invoke, builds the URL (substituting path params + appending the query string), attaches
Authorization: Bearer ${OPENDATA_API_KEY}for non-GET, sends the request, and returns the response as a text MCPcontentblock.
The auth contract is enforced at the client layer (src/api-client.ts):
- GET → no auth header is sent. If you pass a key it's still ignored.
- Non-GET →
OPENDATA_API_KEYis required; the request fails fast with a clear error if it's missing.
Logging, retries, and shutdown
- Structured logging to stderr (stdout is reserved for the MCP protocol). One JSON object per line:
{ts, level, msg, ...fields}. Tuned viaOPENDATA_LOG_LEVEL(off|error|info|debug, defaultinfo). Every tool invocation carries a shortrid(request id) that also surfaces in any error message returned to the client — so a user can quote anERROR (rid=…)and you can grep it in logs. - Bounded retries on transient failures only:
502 / 503 / 504and network errors are retried with jittered exponential backoff (200ms → 600ms → 1.4s, capped).500and any4xxare never retried — a 500 means "I tried, it broke" and retrying just hammers a server that already gave its answer; a 4xx means the request is wrong and retrying won't help. Total attempts viaOPENDATA_MAX_ATTEMPTS(default 3). - Graceful shutdown on
SIGINT/SIGTERM: the MCP transport is closed beforeprocess.exitso any in-flight tool calls have a chance to land.
Layout
src/
index.ts # entry — MCP server, stdio transport
openapi-loader.ts # loads + normalizes the bundled spec
openapi-to-zod.ts # OpenAPI schema → Zod schema converter
api-client.ts # fetch wrapper + auth contract
tools.ts # generates MCP tool defs from normalized ops
__tests__/
tools.test.ts # unit tests (mocked fetch)
live.test.ts # live tests (skipped unless OPENDATA_BASE_URL set)
openapi.json # bundled OpenAPI 3.1 spec
License
MIT
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。