Thursday, 19 November 2026
Shola JegedeDeveloper tools
mcp-starter-kit
Production foundation for MCP servers: any auth provider, any database, any billing system, any runtime.
npm install mcp-starter-kit38downloads last week
1,541downloads last 12 months
v2.0.0latest, Nov 19, 2026
33releases since Oct 10, 2025
Downloads
Download history
Loading download history…
Build this
Build with mcp-starter-kit
- 1
Why it exists
Getting a demo MCP server to answer
tools/listtakes ten minutes. Getting one ready for real users takes much longer, because the work that matters sits around the tools:- verifying tokens properly (signature, issuer, audience, expiry) and publishing the OAuth metadata MCP clients look for
- knowing which tenant a call belongs to and passing that to your database layer
- per-tool scopes, rate limits and usage quotas that hold across instances
- asking the user before a destructive action runs
- keeping internal errors away from the model while still giving it messages it can act on
- calling an LLM from inside a tool now that sampling is deprecated
- running the same server over stdio for local clients and over HTTP behind a load balancer
The kit does those things once, with adapters, so the same server code works with Kinde or Auth0 or Entra, with Postgres or Redis or Mongo, on Node or Cloudflare Workers.
- 2
How a tool call flows
codeMCP client (Claude, ChatGPT, Cursor, VS Code, an agent) │ Authorization: Bearer <token> ▼ ┌──────────────────────────── mcp-starter-kit ─────────────────────────────┐ │ auth adapter verify token, map claims to a Principal │ │ (Kinde, Auth0, Clerk, WorkOS, Okta, Entra, Keycloak, │ │ Cognito, Supabase, Firebase, Better Auth, any OIDC, │ │ opaque tokens, API keys) │ │ scope check per tool; OAuth clients get a step-up challenge │ │ confirmation destructive tools ask the user first │ │ rate limit per user, per org or global, in the shared store │ │ entitlements quotas, Stripe, Kinde, Autumn, Polar, or your own │ │ your handler typed input, ctx.services, ctx.llm, ctx.progress │ │ result structured output, safe errors, usage settled │ └──────────────────────────────────────────────────────────────────────────┘ │ │ │ store adapter LLM router runtime memory, Postgres, OpenRouter, Vercel stdio, Node HTTP, Workers, SQLite, MySQL, AI Gateway, LiteLLM, Bun, Deno, Vercel, any Redis, Upstash, Portkey, Cloudflare, fetch-based host Mongo OpenAI, Anthropic