mcp-use

Product Introduction

1. mcp-use is an open-source developer toolkit and cloud infrastructure platform designed to accelerate the creation and deployment of custom AI agents using MCP servers. It provides SDKs for Python and JavaScript with prebuilt integrations for model orchestration, server management, and agent workflows. The platform has been adopted by over 100k developers and organizations like NASA and NVIDIA for AI infrastructure automation.
2. The core value lies in its ability to unify MCP server management through a single API endpoint while maintaining compatibility with multiple deployment environments. It eliminates infrastructure complexity through zero-friction server pooling, automatic tool discovery, and built-in observability features for production AI systems.

Main Features

1. The SDK enables 3-line code implementation of AI agents with native support for multi-step tool execution, demonstrated through the MCPAgent class that integrates LLM reasoning (Claude 3.5 Sonnet), tool selection, and server communication. Developers can chain operations like MCPClient(server_pool="marketing_servers") to aggregate distributed resources.
2. MCP Gateway offers enterprise-grade routing with automatic load balancing across hosted servers, sandboxed VM instances, and third-party services. The gateway handles OAuth, ACLs, and real-time metrics monitoring through a unified control plane accessible via web dashboard or API.
3. Server lifecycle management supports three deployment modes: ephemeral stdio servers executed via Python/node commands (e.g., stdio_client(server_params)), persistent cloud-hosted instances with auto-scaling, and proxy configurations for external services. All modes feature automatic tool schema discovery through the session.list_tools() method.

Problems Solved

1. Eliminates manual orchestration of multiple MCP server instances by providing dynamic server pooling and tool aggregation. The server_pool parameter in MCPClient automatically routes requests to available resources while maintaining session state through AsyncExitStack management.
2. Targets engineering teams building production AI systems that require integration of diverse tools (Google Drive, Slack, custom APIs) with LLM-powered agents. The platform is particularly valuable for enterprises needing audit trails, shown through built-in metrics and tracing capabilities in the gateway.
3. Enables rapid prototyping of AI workflows like content generation (agent.run("Create post about release")) and data processing pipelines. Use cases include automated customer support systems, multi-tool research assistants, and real-time monitoring agents interacting with MCP-enabled services.

Unique Advantages

1. Unlike standalone AI orchestration frameworks, mcp-use combines SDK flexibility with managed infrastructure through its hybrid architecture. The stdio server implementation (StdioServerParameters) allows local debugging while maintaining compatibility with cloud deployment patterns.
2. Introduces tool chaining with automatic schema validation, where the session.call_tool() method enforces input/output contracts using JSON Schema from server-defined input_schema metadata. This prevents runtime errors in multi-step agent operations.
3. Offers production advantages through the managed gateway's distributed tracing system, which correlates agent activities across server boundaries. The platform's 1-click deployment for MCP servers outperforms competitors by reducing cold start times through pre-warmed container pools.

Frequently Asked Questions (FAQ)

1. How do I connect to existing MCP servers? Use the MCPClient constructor with either server URLs or predefined pool identifiers like server_pool="marketing_servers", which automatically authenticates through the gateway using environment-based credentials.
2. Can I run MCP servers locally for testing? Yes, execute .py or .js server scripts via stdio_client with StdioServerParameters, which creates isolated processes while enabling full tool discovery through the session.initialize() handshake protocol.
3. How does authentication work for third-party services? The gateway handles OAuth flows and API key management through centralized secrets storage, accessible to agents via role-based access control (RBAC) policies defined in the dashboard.
4. What LLM models are supported? While optimized for Claude 3.5 Sonnet (as shown in anthropic.messages.create()), the SDK allows integration with any model provider through the AnyModelProvider abstraction layer.
5. How is traffic distributed across servers? The gateway employs least-connections load balancing with circuit breakers, automatically rerouting requests from unresponsive servers while maintaining sticky sessions when required by tool state.