strands-tools-mcp

Expose Strands Agents tools as an MCP server via stdio transport, so any MCP client (Kiro, Claude Desktop, etc.) can use them.

Supports both @tool-decorated functions and TOOL_SPEC module-based tools from the strands-agents-tools package, as well as custom tool files.

Installation

pip install strands-tools-mcp

Or from source:

git clone https://github.com/mkmeral/strands-tools-mcp.git
cd strands-tools-mcp
pip install -e ".[dev]"

Configuration

Configure which tools to expose via environment variables:

Both variables can be used together. At least one must be set.

Usage

As a CLI

STRANDS_TOOLS=current_time,shell strands-tools-mcp

In Kiro / Claude Desktop MCP Config

Add to your MCP configuration file:

{
  "mcpServers": {
    "strands-tools": {
      "command": "strands-tools-mcp",
      "env": {
        "STRANDS_TOOLS": "shell,http_request,file_read,file_write"
      }
    }
  }
}

With Custom Tools

You can combine built-in Strands tools with your own custom tool files:

{
  "mcpServers": {
    "my-tools": {
      "command": "strands-tools-mcp",
      "env": {
        "STRANDS_TOOLS": "current_time",
        "STRANDS_TOOLS_PATHS": "/path/to/my_tool.py"
      }
    }
  }
}

With Remote Tools (URLs)

Point directly at raw GitHub links, gists, or any hosted .py file:

{
  "mcpServers": {
    "remote-tools": {
      "command": "strands-tools-mcp",
      "env": {
        "STRANDS_TOOLS_PATHS": "https://raw.githubusercontent.com/user/repo/main/my_tool.py"
      }
    }
  }
}

Local paths and URLs can be mixed in STRANDS_TOOLS_PATHS.

Custom tool files should use the @tool decorator from strands:

from strands import tool

@tool
def my_custom_tool(message: str) -> str:
    """A custom tool that processes a message.

    Args:
        message: The message to process

    Returns:
        The processed message
    """
    return f"Processed: {message}"

How It Works

The server uses the Strands SDK's built-in tool loader to discover tools, then exposes them via the MCP protocol's stdio transport using the raw mcp library (no FastMCP). Both tool patterns are supported:

• @tool decorated functions (DecoratedFunctionTool) — e.g. current_time, shell, calculator
• TOOL_SPEC module-based tools (PythonAgentTool) — e.g. file_read, file_write, http_request, python_repl

All tools are invoked through the unified AgentTool.stream() interface, so both types work identically.

Development

# Install in dev mode
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Lint and format
ruff check src/ tests/
ruff format src/ tests/

License

Apache-2.0