@langchain/mcp-adapters

Converts Model Context Protocol (MCP) servers into LangChain-compatible Tool objects through a standardized adapter pattern. The adapter introspects MCP server capabilities (resources, prompts, tools) and wraps them as LangChain ToolInterface implementations, enabling seamless integration of MCP-exposed functionality into LangChain agent chains without manual schema translation or binding code.

Extracts and injects MCP server resources (documents, files, structured data) into LangChain's context/memory systems through a resource adapter. The adapter reads MCP resource URIs, fetches content via the MCP protocol, and converts them into LangChain-compatible context formats (Document objects, memory stores, or RAG-ready embeddings), enabling agents to access external knowledge without explicit tool calls.

Validates MCP tool results against declared schemas and enforces type safety through a validation layer that parses tool responses, checks against JSON Schema definitions, and raises errors for schema violations. The validator supports custom validation rules, type coercion, and detailed error reporting, preventing downstream errors from malformed MCP responses and enabling type-safe tool result handling in LangChain chains.

Orchestrates multiple MCP servers and routes tool calls to appropriate servers based on capability matching, load balancing, or explicit routing rules through a routing layer. The layer maintains a registry of available MCP servers, their capabilities, and health status, matches incoming tool requests to capable servers, and distributes load across servers, enabling agents to leverage multiple MCP servers transparently without explicit server selection.

Converts MCP prompt definitions (reusable prompt templates with arguments) into LangChain PromptTemplate objects through schema introspection and binding. The adapter reads MCP prompt metadata (name, description, arguments), maps argument types to LangChain variable placeholders, and creates executable prompt templates that can be chained with LLMs, enabling prompt reuse across MCP and LangChain ecosystems.

Provides a unified transport abstraction for MCP communication (stdio, HTTP, WebSocket) that abstracts protocol-level details from LangChain adapters. The layer handles connection lifecycle (setup, teardown, reconnection), message serialization (JSON-RPC 2.0), and error handling, allowing adapters to work with any MCP transport without transport-specific code, enabling flexible deployment (local servers, remote APIs, containerized services).

Implements standardized error handling and exponential backoff retry logic for MCP communication failures through a resilience layer. The layer catches MCP protocol errors (timeouts, connection failures, invalid responses), applies configurable retry strategies (exponential backoff, jitter), and provides detailed error context to LangChain agents, enabling graceful degradation and automatic recovery without explicit error handling in adapter code.

Automatically discovers and introspects MCP server capabilities (available tools, resources, prompts, sampling methods) through protocol-level introspection without requiring manual capability declarations. The discovery mechanism queries the MCP server's capability manifest, parses tool schemas, resource types, and prompt definitions, and exposes them as queryable metadata, enabling dynamic tool registration and capability-aware agent routing.

Integrates MCP sampling methods (server-side LLM sampling capabilities) with LangChain's LLM interface through a sampling adapter. The adapter wraps MCP sampling method definitions, handles sampling parameter marshaling, and executes sampling requests through the MCP protocol, enabling LangChain agents to leverage server-side LLM capabilities (e.g., specialized models, fine-tuned variants) without local LLM instantiation.

Manages context window allocation and token budgeting for MCP-integrated agents through a context manager that tracks token usage across MCP resources, tools, and prompts, and implements strategies for context prioritization (recent messages, high-relevance resources, critical tools). The manager prevents context overflow by truncating or filtering low-priority context, enabling agents to operate within LLM token limits while maintaining access to MCP capabilities.

Caches MCP tool execution results and resource fetches through a memoization layer that stores results keyed by tool name, arguments, and timestamp, enabling agents to reuse results for identical or similar requests without re-executing MCP tools. The cache supports TTL-based expiration, LRU eviction, and optional persistent storage, reducing latency and MCP server load for repeated queries.

Persists MCP-integrated agent state (tool execution history, context, decision traces) to external storage and enables recovery from failures through a state management layer. The layer serializes agent state including MCP tool calls, results, and context, stores it in configurable backends (file system, database, cloud storage), and restores state on recovery, enabling long-running agents to resume from checkpoints without losing progress.