feat(manage_tools): add runtime tool management capabilities

[cagataycali]

Summary

Add manage_tools tool for dynamic tool management at runtime - a package manager for AI agent tools.

Features

• list: Display all currently registered tools
• add: Load tools from packages, files, URLs, or GitHub
• remove: Unregister tools by name
• reload: Hot-reload tools to pick up code changes
• create: Generate new tools from Python code on-the-fly
• fetch: Download and load tools from remote URLs
• discover: List all available tools in a module with introspection
• sandbox: Validate code in isolated subprocess before loading

Key Highlights

• Package manager for AI agent tools - expand capabilities without restarts
• Auto-installation of missing PyPI packages via pip or uv
• Sandbox testing validates code before loading into agent
• GitHub blob URL auto-conversion to raw URLs
• Comprehensive test coverage

Environment Variables

Variable	Description	Default
STRANDS_AUTO_INSTALL_TOOLS	Enable automatic package installation	false
STRANDS_PACKAGE_INSTALLER	Package installer (pip or uv)	pip
STRANDS_INSTALL_TIMEOUT	Installation timeout (seconds)	300
STRANDS_DISABLE_LOAD_TOOL	Disable dynamic tool loading	false
STRANDS_TOOLS_CACHE_DIR	Cache directory for fetched tools	temp dir

Usage Examples

from strands import Agent
from strands_tools import manage_tools

agent = Agent(tools=[manage_tools])

# List registered tools
agent.tool.manage_tools(action="list")

# Add a tool from a package
agent.tool.manage_tools(action="add", tools="strands_tools.calculator")

# Add with auto-installation
agent.tool.manage_tools(action="add", tools="strands_mlx.dataset_splitter", install=True)

# Discover available tools in a module
agent.tool.manage_tools(action="discover", tools="strands_tools", verbose=True)

# Create a custom tool at runtime
agent.tool.manage_tools(
    action="create",
    code='''
    from strands import tool

    @tool
    def multiply(a: int, b: int) -> int:
        """Multiply two integers."""
        return a * b
    '''
)

# Fetch from GitHub
agent.tool.manage_tools(
    action="fetch",
    url="https://github.com/user/repo/blob/main/my_tool.py"
)

# Sandbox test code before loading
agent.tool.manage_tools(action="sandbox", code="from strands import tool\n...")

# Remove tools
agent.tool.manage_tools(action="remove", tools="calculator")

# Hot-reload tools
agent.tool.manage_tools(action="reload", tools="shell")

Testing

• Comprehensive test suite with high coverage
• Tests cover all actions, helper functions, and edge cases
• Sandbox testing validation
• Package installation mocking

Checklist

• Code follows project style guidelines
• Tests pass locally
• Documentation included in docstrings
• No breaking changes

[JackYPCOnline]

we had some secruity issues related to url, we need to restrict urls, do validations.

[cagataycali]

network limitations should be applied on runtime level, not on tool to be secure. for example, shell tool have unrestricted access to shell but OS layer determine that operation is successful or not, same applies for this tool

[mkmeral]

/strands review