Build MCP Server Sdk V2

Build MCP Server (SDK v2 Alpha)

Build and maintain MCP servers on the v2 alpha split-package SDK: @modelcontextprotocol/server, @modelcontextprotocol/client, @modelcontextprotocol/core, plus /node, /express, /hono adapters. ESM-only, Node 20+, Zod v4. Status as of 2026-05-09: latest npm tag is 2.0.0-alpha.2 — pin exact, plan rollback.

When to use

Trigger this skill if any of these are true:

• Building a brand-new MCP server and the user picks v2, "the alpha", or split packages.
• package.json already depends on @modelcontextprotocol/server, @modelcontextprotocol/client, or @modelcontextprotocol/core.
• Existing code uses new McpServer(...) from @modelcontextprotocol/server and server.registerTool(...) with the high-level API.
• Tool/resource/prompt handlers use (args, ctx) with ctx.mcpReq.signal, ctx.mcpReq.log(), ctx.mcpReq.notify(), or ctx.http?.authInfo.
• HTTP work uses NodeStreamableHTTPServerTransport from @modelcontextprotocol/node, or createMcpExpressApp() / createMcpHonoApp() from the official adapters.
• Schemas are full z.object({...}) from zod/v4, not raw-shape shorthand.

Do NOT use this skill if any of these are true:

• package.json depends on the single-package @modelcontextprotocol/sdk (v1) — use build-mcp-server-sdk-v1 instead.
• *Handlers use (args, extra) with extra.sendNotification, extra.authInfo, or extra.signal — that is v1; use build-mcp-server-sdk-v1.*
• The job is porting an existing v1 server to v2 — use convert-mcp-sdk-v1-to-v2 (covers package split, import rewrite, extra → ctx mapping, OAuth replacement, staging strategy).
• *The project uses the mcp-use wrapper or @hono/mcp community middleware — use build-mcp-use-server, or migrate before applying official adapter patterns.*
• The user wants an agentic-quality / hardening / context-budget audit, not SDK correctness — pair this skill with the relevant build-mcp-* reference for protocol patterns.

Detect v2 vs v1

Run tree -L 3 and read package.json. v2 fingerprints (any one is sufficient):

v1 anti-fingerprints (treat as wrong skill, redirect):

• @modelcontextprotocol/sdk single package → build-mcp-server-sdk-v1
• extra.sendNotification, extra.authInfo, extra.signal → build-mcp-server-sdk-v1
• SSEServerTransport → v1 only; v2 removed it

Core rules

• Always use McpServer from @modelcontextprotocol/server. The low-level Server class is deprecated for direct use.
• Always use registerTool / registerResource / registerPrompt. Positional overloads were removed in v2.
• Always pass full Zod v4 schemas (z.object({...})). Raw shapes are a v1 pattern; if a current alpha still accepts them, treat that as a migration shim, not the target.
• Always import HTTP transport from @modelcontextprotocol/node (e.g. NodeStreamableHTTPServerTransport). SSEServerTransport is removed.
• For Express, use @modelcontextprotocol/express (createMcpExpressApp()). For Hono, use @modelcontextprotocol/hono. Do not silently substitute the community @hono/mcp package.
• Server-side OAuth is removed from the SDK. Wire authentication at the HTTP layer (Passport, custom Bearer middleware, jose) and forward auth into ctx.http?.authInfo. Treat any @modelcontextprotocol/server-auth-legacy as planned/open until npm publish is confirmed.
• ESM-only. No CommonJS dual-publish. Node.js 20+ required.
• Pin alpha versions exactly (--save-exact); never use ^ ranges across alphas.

Workflow

1 — Detect what exists

Inspect package.json and src/. Decide: existing v2 server (go to 2A), new v2 server (go to 2B), or wrong skill (redirect per When to use and stop).

2A — Maintain or fix an existing v2 server

Read the implementation. Verify:

• Context usage: ctx.mcpReq.signal, ctx.mcpReq.log(), ctx.mcpReq.notify(), ctx.http?.authInfo. Flag any extra.* access — that is v1 leakage.
• Schemas: full z.object() (not raw shapes) for new code. outputSchema present whenever the tool returns structuredContent.
• Transport: NodeStreamableHTTPServerTransport from @modelcontextprotocol/node for HTTP; StdioServerTransport from @modelcontextprotocol/server for stdio.
• Framework: createMcpExpressApp() or createMcpHonoApp() for HTTP framework wiring (DNS rebinding protection lives in the adapter).
• Annotations: readOnlyHint, destructiveHint, idempotentHint, openWorldHint set deliberately for tools with side effects.

Then make the requested change (add tool, fix bug, add auth middleware, etc.).

2B — Scope a new v2 server

Decide:

1. Wraps what? API, database, filesystem, CLI, or in-process logic.
2. Transport? stdio for local; Streamable HTTP for remote/multi-client.
3. Framework? Express or Hono if HTTP — both have first-party adapters.
4. Auth? External AS + middleware; SDK no longer hosts an authorization server.

3 — Choose the implementation branch

4 — Preflight setup

• Node.js 20+ installed
• If existing: run bash scripts/check-mcp-server-v2-version.sh from the project root (see scripts/check-mcp-server-v2-version.sh.md); unsafe alpha ranges must fail
• npm install --save-exact @modelcontextprotocol/server@2.0.0-alpha.2
• npm install zod@^4
• HTTP also: npm install --save-exact @modelcontextprotocol/node@2.0.0-alpha.2
• Express also: npm install --save-exact @modelcontextprotocol/express@2.0.0-alpha.2 express
• Hono also: npm install --save-exact @modelcontextprotocol/hono@2.0.0-alpha.2 hono
• "type": "module" in package.json
• TypeScript 5+, "module": "Node16", "moduleResolution": "Node16"

5 — Build

Default sequence:

1. Construct McpServer with { name, version } and optional { instructions, capabilities }.
2. Define Zod v4 schemas: z.object({ field: z.string() }) (full schemas, not raw shapes).
3. Register tools with server.registerTool(name, config, handler) — inputSchema, annotations, handler (args, ctx) => CallToolResult.
4. Register resources with server.registerResource() if exposing data.
5. Register prompts with server.registerPrompt() if providing templates.
6. Construct transport, then await server.connect(transport).
7. Handle graceful shutdown (SIGINT/SIGTERM → await server.close()).

6 — Validate

• Local checks first: npm run build, focused tests if present.
• stdio: npx @anthropic-ai/mcp-inspector npx tsx src/index.ts.
• HTTP: start server; probe with curl or Inspector.
• Live CLI smoke: if mcpc is installed, hand off to test-by-mcpc-cli. Minimum sequence: initialize → tools/list → one successful call → one invalid-arg call returning isError: true.
• Schemas: invalid input → tool error (isError: true), not a thrown protocol error.
• Context: confirm ctx.mcpReq is the access path, never extra.

Quick start — minimal v2 stdio server

import { McpServer, StdioServerTransport } from "@modelcontextprotocol/server";
import * as z from "zod/v4";

const server = new McpServer(
  { name: "my-server", version: "1.0.0" },
  { instructions: "A helpful server" }
);

server.registerTool("greet", {
  title: "Greet User",
  description: "Greet a user by name",
  inputSchema: z.object({ name: z.string().describe("The user's name") }),
  annotations: { readOnlyHint: true, destructiveHint: false },
}, async ({ name }, ctx) => {
  await ctx.mcpReq.log("info", `Greeting ${name}`);
  return { content: [{ type: "text" as const, text: `Hello, ${name}!` }] };
});

const transport = new StdioServerTransport();
await server.connect(transport);

Core API summary

McpServer

new McpServer(
  { name: string, version: string, description?: string, icons?: Icon[] },
  { capabilities?: ServerCapabilities, instructions?: string }
)

server.connect(transport: Transport): Promise<void>
server.close(): Promise<void>
server.registerTool(name, config, handler): RegisteredTool
server.registerResource(name, uri | template, config, handler): RegisteredResource
server.registerPrompt(name, config, handler): RegisteredPrompt
server.sendToolListChanged(): void
server.sendResourceListChanged(): void
server.sendPromptListChanged(): void
server.sendLoggingMessage(params): Promise<void>
server.isConnected(): boolean
server.experimental.tasks  // ExperimentalMcpServerTasks

registerTool config

{
  title?: string,
  description?: string,
  inputSchema?: AnySchema,           // z.object({...}) — full Zod v4 schema
  outputSchema?: AnySchema,          // enables structuredContent validation
  annotations?: ToolAnnotations,
  _meta?: Record<string, unknown>,
}

ServerContext (handler second argument)

// Tool handler: (args, ctx) => CallToolResult
// No-arg tool:  (ctx)       => CallToolResult

ctx.sessionId?: string
ctx.mcpReq.id: RequestId
ctx.mcpReq.method: string
ctx.mcpReq.signal: AbortSignal
ctx.mcpReq._meta?: RequestMeta
ctx.mcpReq.send(request, schema, options?): Promise<Result>
ctx.mcpReq.notify(notification): Promise<void>
ctx.mcpReq.log(level, data, logger?): Promise<void>
ctx.mcpReq.elicitInput(params): Promise<ElicitResult>
ctx.mcpReq.requestSampling(params): Promise<CreateMessageResult>
ctx.http?.authInfo?: AuthInfo
ctx.http?.req?: RequestInfo
ctx.http?.closeSSE?(): void
ctx.http?.closeStandaloneSSE?(): void
ctx.task?.id?: string
ctx.task?.store?: RequestTaskStore

Error handling

import { ProtocolError, ProtocolErrorCode } from "@modelcontextprotocol/core";

// Hard protocol errors:
throw new ProtocolError(ProtocolErrorCode.InvalidParams, "Bad input");

// Soft tool errors (LLM can self-correct):
return { content: [{ type: "text", text: "Error: not found" }], isError: true };

Decision rules

• Use full z.object({...}) for every new tool schema. Raw shapes are v1 style; even if accepted, do not target them.
• Prefer isError: true for recoverable failures — the LLM self-corrects from soft errors.
• Prefer ctx.mcpReq.log() over console.error() so logs reach the client.
• Prefer ctx.mcpReq.elicitInput() over hand-rolled ctx.mcpReq.send() for user input requests.
• Use createMcpExpressApp() / createMcpHonoApp() instead of raw Express/Hono setup — DNS rebinding is handled inside.
• Set every relevant annotations field deliberately; fill all four when safety or side-effects matter.

Guardrails

• Never write new v2-native code with raw Zod shapes — always full z.object().
• Never use extra.sendNotification / extra.authInfo / extra.signal — those are v1; the v2 access path is ctx.mcpReq.* and ctx.http?.authInfo.
• Never import from @modelcontextprotocol/sdk — that is the v1 single package; in v2 you import from /server, /client, /core, /node, /express, /hono.
• Never use SSEServerTransport — removed in v2; use Streamable HTTP.
• Never implement server-side OAuth in the SDK — removed in v2; integrate at the HTTP layer.
• Never use CommonJS — v2 is ESM-only.
• Never run on Node < 20.
• Never use ^ ranges for alpha packages — pin exact and plan rollback.

Compatibility and adoption note

v2 is pre-release alpha as of 2026-05-09. The latest npm split packages are at 2.0.0-alpha.2; main-branch PRs labeled v2.0.0-bc may not yet be published. Most production servers should remain on v1.x until v2 cuts a non-alpha stable release.

In practice:

• Pin alpha versions exactly (no ^); alphas can break between patches.
• Plan rollback before deploying — keep the v1 branch deployable.
• The @modelcontextprotocol/sdk meta-package remains v1 on npm unless fresh npm view proves otherwise.
• @modelcontextprotocol/server-auth-legacy is planned/open; treat it as unpublished until npm view succeeds.
• Verify each MCP host (Claude Desktop, Cursor, Cline, custom) end-to-end on v2 features before depending on them.

Output contract

Report v2 server work with:

1. Target path and detected channel/version.
2. Transport (stdio, Streamable HTTP) and framework (none, Express, Hono).
3. Tools, resources, and prompts added or changed.
4. Auth shape (none, Bearer middleware, Passport, jose, external AS).
5. Validation rung reached and exact commands run.
6. Alpha-risk caveats and rollback status.