PHP MCP Server

Automatically discovers and registers MCP elements (Tools, Resources, Prompts, Resource Templates) by scanning the filesystem for PHP classes annotated with #[McpTool], #[McpResource], #[McpResourceTemplate], and #[McpPrompt] attributes. The Discoverer component uses reflection to parse these attributes and automatically register handlers without manual configuration, enabling zero-boilerplate exposure of application functionality to AI assistants.

Generates JSON Schema 2020-12 compliant schemas automatically by parsing PHP method signatures, type hints, and docblock annotations using reflection and docblock parsing. This eliminates manual schema definition while supporting complex types (unions, generics, nullable types) and docstring-based parameter descriptions, enabling AI clients to understand tool capabilities without separate schema files.

Implements the JSON-RPC 2.0 specification for message exchange between client and server. The Protocol component parses incoming JSON-RPC requests, routes them to appropriate handlers through the Dispatcher, and formats responses according to JSON-RPC 2.0 spec (including error responses with error codes and messages). Supports both request/response and notification patterns, enabling bidirectional communication between MCP clients and servers.

Provides a fluent, chainable API for configuring the MCP server through the ServerBuilder class. Developers use method chaining to register transports, set up dependency injection, configure caching, enable session management, and register MCP elements. The builder pattern enables readable, self-documenting server configuration that can be version-controlled and easily modified without touching core server logic.

Implements StreamableHttpServerTransport for production deployments, supporting resumable connections and event sourcing patterns. Clients can reconnect and resume interrupted streams without losing messages, and the server can emit events as Server-Sent Events (SSE) or streaming JSON responses. This transport is recommended over deprecated HttpServerTransport for new projects requiring reliable message delivery and connection resilience.

Abstracts network communication through pluggable transport implementations (StdioServerTransport, HttpServerTransport, StreamableHttpServerTransport) that all conform to a common interface. The Protocol component handles JSON-RPC 2.0 message parsing and routing independently of transport, allowing the same server logic to operate over STDIO, HTTP+SSE, or streaming HTTP without code changes.

Integrates with PSR-11 Container interface to enable dependency injection for MCP element handlers. The ServerBuilder and Dispatcher automatically resolve handler dependencies from the container, allowing handlers to declare constructor dependencies that are automatically injected without manual wiring. Supports both explicit container configuration and automatic resolution of registered services.

Provides SessionManager component supporting multiple storage backends (in-memory, file-based, Redis, database) for maintaining client session state across requests. Implements automatic garbage collection of expired sessions and supports configurable TTL per session, enabling stateful MCP interactions where clients can maintain context across multiple tool invocations without re-sending full context.

Built on ReactPHP's event loop to enable non-blocking, asynchronous request handling with high concurrency. Components communicate through events rather than direct method calls, allowing the server to handle multiple concurrent client connections without blocking. The event-driven design enables long-running operations (database queries, API calls) to execute concurrently without spawning threads or processes.

Executes MCP tools by routing requests through the Dispatcher to registered tool handlers, validating input against the auto-generated JSON schema before execution. Implements structured error handling that captures exceptions and returns them as JSON-RPC error responses, enabling AI clients to understand and recover from tool execution failures without server crashes.

Serves static resources and dynamic resource templates through a URI-based access pattern. Resources are static content (files, strings) accessible via fixed URIs, while Resource Templates use URI patterns with variable substitution (e.g., `file://{path}`) to dynamically generate content. The Registry stores both types and the Dispatcher routes requests based on URI matching, enabling AI clients to access application data through a standardized resource interface.

Defines and serves prompt templates that generate conversation messages for AI clients. Prompts are registered with templates containing placeholders that are filled with arguments at runtime, enabling the server to provide context-specific conversation starters or system prompts to AI assistants. The Registry stores prompt definitions and the Dispatcher routes prompt requests, allowing AI clients to discover and use server-provided prompts.

Integrates with PSR-16 Cache interface to enable caching of frequently accessed data (schemas, resources, prompts, session data). The Registry and other components can use the cache to avoid expensive operations (reflection, file I/O, database queries). Supports multiple cache backends (in-memory, file, Redis, APCu) through the PSR-16 interface, with configurable TTL per cache entry.