mcp-production-kit
A production-ready MCP server template with OAuth 2.1, RBAC, and audit logging for building secure, observable tool servers.
README
mcp-production-kit
Production-ready TypeScript MCP server template — OAuth 2.1, RBAC, audit logs.
Most MCP server examples stop at "it works in the Inspector". Production needs the other 80%: who is calling, what are they allowed to do, and what did they actually do. This is a clone-and-adapt remote MCP server that ships that part:
- Streamable HTTP transport via the official
@modelcontextprotocol/sdk - OAuth 2.1 resource server — token validation (JWKS), protected-resource metadata, audience binding, per the current MCP authorization spec
- Per-tool RBAC — deny-by-default scope map; read-only tools by default, write tools behind explicit scopes + confirmation payloads
- Append-only audit logging — actor, tool, args hash, status, latency, trace id for every call; GDPR-aware (args hashed, PII redaction hooks)
- Rate limiting, tests, evals — vitest + scripted MCP Inspector flows
- EU-first deployment — Hetzner runbook (Fly.io alternative), Keycloak as the default self-hosted IdP
What this is not
Not a gateway, not a multi-tenant control plane, not a hosted service, not a tool marketplace. One server, one SaaS, done well.
Status
Early — M1 of the roadmap is done: the Streamable HTTP
server runs, every tool call is audited (stdout JSON + Postgres sink), health
endpoints and the Hetzner deploy path are in place. OAuth 2.1 + RBAC against
a real IdP (M2) and the runnable example (M3) are next — until M2 lands,
scopes come from the DEV_GRANTED_SCOPES dev-only escape hatch, so don't
point this at production data yet.
Layout
src/server.ts Streamable HTTP transport, session handling (M1)
src/tools/ tool registry — zod schemas, task-oriented tools
src/auth/ OAuth 2.1 resource server (M2)
src/rbac/ per-tool scope map, deny-by-default
src/audit/ structured audit middleware, append-only sink
src/rate-limit/ per-client + per-tool limits (M3)
src/observability/ OpenTelemetry, health endpoints (M1)
adapters/ IdP adapters: Keycloak (default), Auth0, WorkOS
example/ Supabase demo SaaS, 5 tools (M3)
test/ vitest + MCP Inspector scripts + tool evals
deploy/ Dockerfile, compose, Hetzner runbook (M1)
docs/ tool-design guide, adaptation guide, SECURITY.md
Development
Node ≥ 22, pnpm.
pnpm install
pnpm lint && pnpm test && pnpm build
Design principles live in docs/tool-design.md — start there before adding tools.
License
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。