What it does
MCPAdapt provides a unified adapter layer between Model Context Protocol servers and agentic frameworks. It translates tools from any of the 650+ MCP servers (cataloged on glama.ai or smithery.ai) into native tool formats for LangChain, CrewAI, Smolagents, Google GenAI, and other supported frameworks. This eliminates framework-specific rewrites: define an MCP server once, and use it across multiple agents. Both local stdio-based servers (invoked as subprocesses) and remote SSE-based servers are supported.
Who it's for
Agent engineers building with LangChain, CrewAI, Smolagents, or Google GenAI who need access to tools or data available as MCP servers but not as native framework plugins. Particularly valuable for teams running multiple frameworks or rapidly prototyping agents that need diverse tool ecosystems.
Common use cases
- Plug a specialized MCP server (filesystem, API, database, search) into a LangChain or CrewAI agent without writing a custom integration
- Expand Smolagents with external tools by registering MCP servers at startup
- Build agents that freely mix native framework tools with MCP ecosystem tools
- Prototype agents that rapidly switch between frameworks while reusing the same MCP tool definitions
Setup pitfalls
- Framework-specific extras are required:
pip install mcpadapt[langchain],mcpadapt[crewai], etc. The base package alone will not work - Security: README warns to only connect to trusted MCP servers, especially over SSE. Remote servers gain broad system and network access
- Adapter mismatch: each framework needs its own adapter class (
SmolAgentsAdapter,LangchainAdapter, etc.); wrong adapter causes runtime errors - Stdio servers require the command and its dependencies in PATH; SSE servers require network connectivity and the remote server running