mcpadapt

Manages bidirectional connections to MCP servers using an adapter pattern that abstracts both StdIO (local subprocess) and SSE (remote HTTP) transport layers. The MCPAdapt class acts as a context manager that establishes connections, negotiates protocol handshakes, maintains connection state, and gracefully closes resources. Supports both synchronous and asynchronous operation patterns through separate code paths, enabling integration with frameworks that require specific concurrency models.

Implements a ToolAdapter interface that defines abstract methods for converting MCP tool specifications (JSON schemas with input/output types) into framework-specific tool formats. Each supported framework (Smolagents, LangChain, CrewAI, Google GenAI) has a concrete adapter that translates MCP's canonical tool schema into that framework's expected tool definition structure, parameter validation rules, and execution signatures. The transformation preserves tool semantics while conforming to each framework's tool calling conventions.

Manages local MCP servers running as subprocesses using the StdIO (standard input/output) transport protocol. MCPAdapt spawns the server process, establishes bidirectional communication through stdin/stdout pipes, and handles process lifecycle events (startup, shutdown, crashes). The StdIO transport is the standard for local MCP servers, enabling integration with tools like Claude Desktop and local development environments.

Connects to remote MCP servers using the Server-Sent Events (SSE) HTTP transport protocol, enabling integration with cloud-hosted or network-accessible MCP servers. MCPAdapt establishes HTTP connections to the server endpoint, negotiates the MCP protocol over SSE, and maintains the connection for tool invocation. This enables integration with MCP servers that don't run locally, such as cloud services or remote development environments.

Enables connecting to multiple MCP servers simultaneously and aggregating their tool catalogs into a unified tool registry. The MCPAdapt class maintains a list of server connections and merges tool definitions from all servers, with built-in deduplication logic to handle tools with identical names across different servers. Tools are exposed as a flat list to the target framework, allowing agents to discover and invoke tools from any connected server without explicit server selection.

Provides dual code paths for synchronous and asynchronous execution, allowing MCPAdapt to integrate with frameworks that have different concurrency requirements. The library exposes both sync context managers and async context managers (mcptools), and framework adapters implement sync/async variants based on framework capabilities. This enables the same MCP server connections to be used in blocking (Smolagents, CrewAI) or non-blocking (LangChain, Google GenAI) frameworks without code duplication.

Executes tool calls by dispatching Remote Procedure Calls (RPCs) to the connected MCP server using the tool name and input parameters. When a framework invokes a tool, MCPAdapt marshals the parameters into the MCP protocol format, sends the call to the server, waits for the response, and returns the result to the framework. This decouples tool execution from the agent framework — the agent doesn't need to know whether tools are implemented locally or remotely on the MCP server.

Resolves JSON schema $ref pointers in MCP tool definitions using the jsonref library, enabling tools to use modular schema definitions with shared type definitions. Validates tool input parameters against the resolved schema before execution, catching type mismatches and missing required fields early. This ensures that tools receive well-formed inputs and that schema references don't cause runtime failures when tools are invoked.

Adapts MCP tools to Smolagents' tool format by implementing the SmolAgentsAdapter class that converts MCP tool schemas into Smolagents Tool objects with proper input/output type annotations. Smolagents tools are synchronous callables that receive parsed input parameters and return results as strings or structured data. The adapter handles the impedance mismatch between MCP's JSON schema format and Smolagents' Python type annotation system.

Adapts MCP tools to LangChain's BaseTool interface by implementing the LangchainAdapter class that converts MCP tool schemas into LangChain Tool objects with proper parameter binding and execution. LangChain tools support both synchronous (_run) and asynchronous (_arun) execution paths, enabling non-blocking tool invocation in async agents. The adapter maps MCP's JSON schema to LangChain's Pydantic-based tool parameter models.

Adapts MCP tools to CrewAI's tool format by implementing the CrewAIAdapter class that converts MCP tool schemas into CrewAI-compatible tool objects. CrewAI tools are assigned to agents with specific roles, enabling role-based access control and tool usage tracking. The adapter integrates with CrewAI's tool registry and execution model, allowing tools to be invoked within CrewAI's multi-agent orchestration framework.

Adapts MCP tools to Google GenAI's function calling format by implementing the GoogleGenAIAdapter class that converts MCP tool schemas into Google's function calling schema. Google GenAI models can invoke functions defined in this format, enabling tool use within Google's generative AI models. The adapter supports both synchronous and asynchronous execution, integrating with Google's async client libraries.