Skip to content

CLAUDE.md

Latest commit

 

History

History
266 lines (188 loc) · 11.5 KB

CLAUDE.md

File metadata and controls

266 lines (188 loc) · 11.5 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Build & Test Commands

pnpm install         # Install all workspace dependencies

pnpm build:all       # Build all packages
pnpm lint:all        # Run ESLint + Prettier checks across all packages
pnpm lint:fix:all    # Auto-fix lint and formatting issues across all packages
pnpm typecheck:all   # Type-check all packages
pnpm test:all        # Run all tests (vitest) across all packages
pnpm check:all       # typecheck + lint across all packages

# Run a single package script (examples)
# Run a single package script from the repo root with pnpm filter
pnpm --filter @modelcontextprotocol/core test                # vitest run (core)
pnpm --filter @modelcontextprotocol/core test:watch          # vitest (watch)
pnpm --filter @modelcontextprotocol/core test -- path/to/file.test.ts
pnpm --filter @modelcontextprotocol/core test -- -t "test name"

Breaking Changes

When making breaking changes, document them in both:

  • docs/migration.md — human-readable guide with before/after code examples
  • docs/migration-SKILL.md — LLM-optimized mapping tables for mechanical migration

Include what changed, why, and how to migrate. Search for related sections and group related changes together rather than adding new standalone sections.

Code Style Guidelines

  • TypeScript: Strict type checking, ES modules, explicit return types
  • Naming: PascalCase for classes/types, camelCase for functions/variables
  • Files: Lowercase with hyphens, test files with .test.ts suffix
  • Imports: ES module style, include .js extension, group imports logically
  • Formatting: 2-space indentation, semicolons required, single quotes preferred
  • Testing: Co-locate tests with source files, use descriptive test names
  • Comments: JSDoc for public APIs, inline comments for complex logic

JSDoc @example Code Snippets

JSDoc @example tags should pull type-checked code from companion .examples.ts files (e.g., client.ts → client.examples.ts). Use ```ts source="./file.examples.ts#regionName" fences referencing //#region regionName blocks; region names follow exportedName_variant or ClassName_methodName_variant pattern (e.g., applyMiddlewares_basicUsage, Client_connect_basicUsage). For whole-file inclusion (any file type), omit the #regionName.

Run pnpm sync:snippets to sync example content into JSDoc comments and markdown files.

Architecture Overview

Core Layers

The SDK is organized into three main layers:

  1. Types Layer (packages/core/src/types/types.ts) - Protocol types generated from the MCP specification. All JSON-RPC message types, schemas, and protocol constants are defined here using Zod v4.

  2. Protocol Layer (packages/core/src/shared/protocol.ts) - The abstract Protocol class that handles JSON-RPC message routing, request/response correlation, capability negotiation, and transport management. Both Client and Server extend this class.

  3. High-Level APIs:

    • Client (packages/client/src/client/client.ts) - Client implementation extending Protocol with typed methods for MCP operations
    • Server (packages/server/src/server/server.ts) - Server implementation extending Protocol with request handler registration
    • McpServer (packages/server/src/server/mcp.ts) - High-level server API with simplified resource/tool/prompt registration

Transport System

Transports (packages/core/src/shared/transport.ts) provide the communication layer:

  • Streamable HTTP (packages/server/src/server/streamableHttp.ts, packages/client/src/client/streamableHttp.ts) - Recommended transport for remote servers, supports SSE for streaming
  • SSE (packages/server/src/server/sse.ts, packages/client/src/client/sse.ts) - Legacy HTTP+SSE transport for backwards compatibility
  • stdio (packages/server/src/server/stdio.ts, packages/client/src/client/stdio.ts) - For local process-spawned integrations

Server-Side Features

  • Tools/Resources/Prompts: Registered via McpServer.tool(), .resource(), .prompt() methods
  • OAuth/Auth: Full OAuth 2.0 server implementation in packages/server/src/server/auth/
  • Completions: Auto-completion support via packages/server/src/server/completable.ts

Client-Side Features

  • Auth: OAuth client support in packages/client/src/client/auth.ts and packages/client/src/client/auth-extensions.ts
  • Client middleware: Request middleware in packages/client/src/client/middleware.ts (unrelated to the framework adapter packages below)
  • Sampling: Clients can handle sampling/createMessage requests from servers (LLM completions)
  • Elicitation: Clients can handle elicitation/create requests for user input (form or URL mode)
  • Roots: Clients can expose filesystem roots to servers via roots/list

Middleware packages (framework/runtime adapters)

The repo also ships “middleware” packages under packages/middleware/ (e.g. @modelcontextprotocol/express, @modelcontextprotocol/hono, @modelcontextprotocol/node). These are thin integration layers for specific frameworks/runtimes and should not add new MCP functionality.

Experimental Features

Located in packages/*/src/experimental/:

  • Tasks: Long-running task support with polling/resumption (packages/core/src/experimental/tasks/)

Zod Schemas

The SDK uses zod/v4 internally. Schema utilities live in:

  • packages/core/src/util/schema.ts - AnySchema alias and helpers for inspecting Zod objects

Validation

Pluggable JSON Schema validation (packages/core/src/validators/):

  • ajvProvider.ts - Default Ajv-based validator
  • cfWorkerProvider.ts - Cloudflare Workers-compatible alternative

Examples

Runnable examples in examples/:

  • examples/server/src/ - Various server configurations (stateful, stateless, OAuth, etc.)
  • examples/client/src/ - Client examples (basic, OAuth, parallel calls, etc.)
  • examples/shared/src/ - Shared utilities (OAuth demo provider, etc.)

Message Flow (Bidirectional Protocol)

MCP is bidirectional: both client and server can send requests. Understanding this flow is essential when implementing new request types.

Class Hierarchy

Protocol (abstract base)
├── Client (packages/client/src/client/client.ts)     - can send requests TO server, handle requests FROM server
└── Server (packages/server/src/server/server.ts)     - can send requests TO client, handle requests FROM client
    └── McpServer (packages/server/src/server/mcp.ts) - high-level wrapper around Server

Outbound Flow: Sending Requests

When code calls client.callTool() or server.createMessage():

  1. High-level method (e.g., Client.callTool()) calls this.request()
  2. Protocol.request():
    • Assigns unique message ID
    • Checks capabilities via assertCapabilityForMethod() (abstract, implemented by Client/Server)
    • Creates response handler promise
    • Calls transport.send() with JSON-RPC request
    • Waits for response handler to resolve
  3. Transport serializes and sends over wire (HTTP, stdio, etc.)
  4. Protocol._onresponse() resolves the promise when response arrives

Inbound Flow: Handling Requests

When a request arrives from the remote side:

  1. Transport receives message, calls transport.onmessage()
  2. Protocol.connect() routes to _onrequest(), _onresponse(), or _onnotification()
  3. Protocol._onrequest():
    • Looks up handler in _requestHandlers map (keyed by method name)
    • Creates BaseContext with signal, sessionId, sendNotification, sendRequest, etc.
    • Calls buildContext() to let subclasses enrich the context (e.g., Server adds requestInfo)
    • Invokes handler, sends JSON-RPC response back via transport
  4. Handler was registered via setRequestHandler('method', handler)

Handler Registration

// In Client (for server→client requests like sampling, elicitation)
client.setRequestHandler('sampling/createMessage', async (request, ctx) => {
  // Handle sampling request from server
  return { role: "assistant", content: {...}, model: "..." };
});

// In Server (for client→server requests like tools/call)
server.setRequestHandler('tools/call', async (request, ctx) => {
  // Handle tool call from client
  return { content: [...] };
});

Request Handler Context

The ctx parameter in handlers provides a structured context:

BaseContext (common to both Server and Client), fields organized into nested groups:

  • sessionId?: Transport session identifier
  • mcpReq: Request-level concerns
    • id: JSON-RPC message ID
    • method: Request method string (e.g., 'tools/call')
    • _meta?: Request metadata
    • signal: AbortSignal for cancellation
    • send(request, schema, options?): Send related request (for bidirectional flows)
    • notify(notification): Send related notification back
  • http?: HTTP transport info (undefined for stdio)
    • authInfo?: Validated auth token info
  • task?: Task context ({ id?, store, requestedTtl? }) when task storage is configured

ServerContext extends BaseContext.mcpReq and BaseContext.http? via type intersection:

  • mcpReq adds: log(level, data, logger?), elicitInput(params, options?), requestSampling(params, options?)
  • http? adds: req? (HTTP request info), closeSSE?, closeStandaloneSSE?

ClientContext is currently identical to BaseContext.

Capability Checking

Both sides declare capabilities during initialization. The SDK enforces these:

  • Client→Server: Client.assertCapabilityForMethod() checks _serverCapabilities
  • Server→Client: Server.assertCapabilityForMethod() checks _clientCapabilities
  • Handler registration: assertRequestHandlerCapability() validates local capabilities

Adding a New Request Type

  1. Define schema in src/types.ts (request params, result schema)
  2. Add capability to ClientCapabilities or ServerCapabilities in types
  3. Implement sender method in Client or Server class
  4. Add capability check in the appropriate assertCapabilityForMethod()
  5. Register handler on the receiving side with setRequestHandler()
  6. For McpServer: Add high-level wrapper method if needed

Server-Initiated Requests (Sampling, Elicitation)

Server can request actions from client (requires client capability):

// Server sends sampling request to client
const result = await server.createMessage({
  messages: [...],
  maxTokens: 100
});

// Client must have registered handler:
client.setRequestHandler('sampling/createMessage', async (request, extra) => {
  // Client-side LLM call
  return { role: "assistant", content: {...} };
});

Key Patterns

Request Handler Registration (Low-Level Server)

server.setRequestHandler('tools/call', async (request, extra) => {
    // extra contains sessionId, authInfo, sendNotification, etc.
    return {
        /* result */
    };
});

Tool Registration (High-Level McpServer)

mcpServer.tool('tool-name', { param: z.string() }, async ({ param }, extra) => {
    return { content: [{ type: 'text', text: 'result' }] };
});

Transport Connection

// Server
// (Node.js IncomingMessage/ServerResponse wrapper; exported by @modelcontextprotocol/node)
const transport = new NodeStreamableHTTPServerTransport({ sessionIdGenerator: () => randomUUID() });
await server.connect(transport);

// Client
const transport = new StreamableHTTPClientTransport(new URL('http://localhost:3000/mcp'));
await client.connect(transport);