@mcp-use/modelcontextprotocol-sdk

Implements the Model Context Protocol server-side runtime using JSON-RPC 2.0 message framing over stdio, WebSocket, or SSE transports. Handles request/response routing, error serialization, and protocol version negotiation through a transport-agnostic abstraction layer that maps incoming MCP messages to TypeScript handler functions.

Provides a declarative API for defining tool schemas (name, description, input parameters) that automatically transpile to OpenAI function-calling format and Anthropic tool_use format. Includes runtime validation of tool invocations against declared schemas using JSON Schema validation, with type-safe TypeScript interfaces generated from schema definitions.

Implements a resource registry that maps URIs (e.g., 'file://path/to/file', 'db://query/users') to content providers. Supports streaming large resources via chunked responses, automatic MIME type detection, and content-type negotiation. Handlers can return text, binary, or structured data with automatic serialization based on declared MIME types.

Provides a registry for storing reusable prompt templates with named placeholders that can be filled at runtime. Supports multi-turn conversation templates with role-based message sequencing (system, user, assistant). Templates are versioned and can reference other templates, enabling composition of complex prompts from simpler building blocks.

Implements a client-side MCP connection handler that manages the lifecycle of connections to MCP servers (stdio, WebSocket, SSE). Automatically handles reconnection with exponential backoff, multiplexes concurrent requests over a single connection, and maintains request/response correlation using JSON-RPC message IDs. Provides a Promise-based API for invoking remote tools and resources.

Implements MCP protocol capability negotiation where servers advertise supported features (tools, resources, prompts) and clients discover available capabilities. Includes version negotiation to ensure client and server compatibility, with fallback mechanisms for older protocol versions. Capabilities are advertised as structured metadata (schemas, descriptions, URIs) that clients can inspect before invoking.

Implements comprehensive error handling that maps application errors to MCP-compliant error codes (InvalidRequest, MethodNotFound, InvalidParams, InternalError, ServerError). Errors are serialized as JSON-RPC 2.0 error objects with detailed messages and optional error data. Includes error context preservation (stack traces, original error objects) for debugging while sanitizing sensitive information in client responses.

Generates TypeScript interfaces and types from MCP tool schemas, resource definitions, and prompt templates. Includes strict type checking for tool arguments, resource URIs, and template variables. Generated types are exported as .d.ts files or inline type definitions, enabling IDE autocomplete and compile-time type validation in handler implementations.

Provides a middleware/hook system that allows intercepting and transforming requests and responses at multiple points in the MCP message pipeline (before/after tool invocation, before/after resource access, before/after error handling). Middleware is executed in order and can modify payloads, add logging, implement caching, or enforce access control. Hooks are async-friendly and support both server-side and client-side middleware.