MCP Agent Tool Adapter

Overview

The MCP Agent Tool Adapter is a bridge that brings the power of external tool execution into AI agents built on the Model Context Protocol (MCP). By converting any MCP‑exposed tool into a first‑class action that an agent can invoke, it removes the friction between language models and real‑world APIs or command‑line utilities. Developers can now build agents that not only reason about user intent but also perform concrete operations—such as querying a database, running shell commands, or calling a REST endpoint—without leaving the MCP ecosystem.

This server exposes two distinct agent backends: a Google ADK implementation that streams responses over FastAPI or a command‑line interface, and a LangGraph implementation that uses the ReAct pattern for dynamic tool planning. The adapter automatically manages MCP server lifecycles, spawning subprocesses as needed, and handles session isolation so that each agent conversation has its own isolated tool context. The result is a plug‑and‑play system where adding a new MCP tool (defined in a simple JSON configuration) instantly unlocks its capabilities for any agent type.

Key features include:

• Modular tool loading – A lightweight loader discovers MCP servers, starts them on demand, and keeps track of active sessions.
• Dual agent support – Switch between Google ADK or LangGraph agents with a single configuration flag, enabling experimentation with different reasoning strategies.
• Streaming and CLI integration – Agents can stream tool outputs to a web UI or terminal, giving developers immediate feedback during development and debugging.
• Extensible adapters – The architecture is designed to accept new MCP tool adapters, making it straightforward to support OpenAPI specifications or custom command‑line tools.

Typical use cases span from building conversational chatbots that can edit files on a server, to creating autonomous agents that navigate complex workflows like data ingestion pipelines or automated testing suites. In each scenario, the adapter abstracts away the boilerplate of tool discovery and session management, allowing developers to focus on agent logic.

By integrating seamlessly with existing MCP servers, the adapter extends AI workflows into real‑world actions while maintaining the stateless, protocol‑driven nature of MCP. Its unique combination of modularity, dual agent support, and easy extensibility makes it a valuable asset for any developer looking to turn language‑model conversations into actionable, tool‑powered interactions.