Tool Manager in LLMling¶

Overview¶

The ToolManager is a central component that manages the lifecycle and availability of tools for agents. It handles: - Tool registration and validation - Tool enabling/disabling - Tool filtering and sorting - MCP server integration - Temporary tool contexts - Tool state management

Core Responsibilities¶

1. Tool Registration and Access¶

from llmling_agent.tools import ToolManager



# Register a simple function as tool
agent.tools.register_tool(
    my_function,
    name_override="custom_name",     # Optional name override
    description_override="Does X",    # Optional description
    enabled=True,                    # Initial state
    source="runtime",                # Tool source
    requires_confirmation=False,     # Whether to confirm execution
    metadata={"custom": "data"}      # Additional metadata
)

# Get available tools
tools = await agent.tools.get_tools(state="enabled")  # filter options
# Get individual tool info
tool_info = manager["tool_name"]  # Dict-like access

2. Temporary Tool Contexts¶

One of the most powerful features is the ability to temporarily modify available tools:

# Temporarily add tools
with agent.tools.temporary_tools([tool1, tool2]) as temp_tools:
    # tool1 and tool2 are available in addition to the agents tools
    ...

# Temporarily add tools and disable all others
with agent.tools.temporary_tools([special_tool], exclusive=True):
    # Only special_tool is available here
    ...

# Original tool states are restored after context exit

Tool Registration Methods¶

Basic Function Registration¶

def my_tool(text: str) -> str:
    """Process some text.

    Args:
        text: Text to process
    """
    return text.upper()

# Register as tool
tool_info = manager.register_tool(my_tool)
tool_info = manager.register_tool("path.to.callable")

@agent.tools.tool
def some_tool(text: str) -> str:
    ...

Worker Agent Registration¶

# Register another agent as a tool
tool_info = manager.register_worker(
    worker_agent,
    name="worker_tool",                  # Optional name override
    reset_history_on_run=True,           # Clear history between runs
    pass_message_history=False,          # Share conversation history
    share_context=False                  # Share agent context
)

MCP Server Integration¶

# Setup MCP server tools
await agent.tools.setup_mcp_servers([
    StdioMCPServer(command="python", args=["-m", "my_server"]),
    SSEMCPServer(url="http://localhost:8000/events")
])

Note: By default the tool manager is already initialzed with MCP server tools from the config when the agent itself enters the async context

Events¶

The manager emits events for tool state changes:

# Tool events
manager.events.added.connect(lambda name, info: print(f"Tool added: {name}"))
manager.events.removed.connect(lambda name: print(f"Tool removed: {name}"))
manager.events.changed.connect(lambda name, info: print(f"Tool changed: {name}"))

# State reset event
manager.tool_states_reset.connect(
    lambda evt: print(f"Tools reset: {evt.previous_tools} -> {evt.new_tools}")
)

Tool Information¶

Each registered tool provides rich metadata through ToolInfo:

tool_info = manager["tool_name"]

print(tool_info.name)                 # Tool name
print(tool_info.description)          # Tool description
print(tool_info.enabled)              # Current state
print(tool_info.source)               # Where tool came from
print(tool_info.requires_confirmation)  # Whether to confirm
print(tool_info.requires_capability)   # Required capability
print(tool_info.metadata)             # Custom metadata
print(tool_info.parameters)           # Parameter information

Capability-Based Tools¶

# Register tool with capability requirement
manager.register_tool(
    sensitive_operation,
    requires_capability="can_execute_code",
    requires_confirmation=True
)