toolbox-core 0.2.0

Python Base SDK for interacting with the Toolbox service

MCP Toolbox Core SDK

This SDK allows you to seamlessly integrate the functionalities of Toolbox allowing you to load and use tools defined in the service as standard Python functions within your GenAI applications.

This simplifies integrating external functionalities (like APIs, databases, or custom logic) managed by the Toolbox into your workflows, especially those involving Large Language Models (LLMs).

• Installation
• Quickstart
• Usage
• Loading Tools
  • Load a toolset
  • Load a single tool
• Invoking Tools
• Synchronous Usage
• Use with LangGraph
• Authenticating to the Toolbox Server
  • When is Client-to-Server Authentication Needed?
  • How it works
  • Configuration
  • Authenticating with Google Cloud Servers
  • Step by Step Guide for Cloud Run
• Authenticating Tools
  • When is Authentication Needed?
  • Supported Authentication Mechanisms
  • Step 1: Configure Tools in Toolbox Service
  • Step 2: Configure SDK Client
    • Provide an ID Token Retriever Function
    • Option A: Add Authentication to a Loaded Tool
    • Option B: Add Authentication While Loading Tools
  • Complete Authentication Example
• Binding Parameter Values
  • Why Bind Parameters?
  • Option A: Binding Parameters to a Loaded Tool
  • Option B: Binding Parameters While Loading Tools
  • Binding Dynamic Values
• Contributing
• License
• Support

Installation

pip install toolbox-core

[!NOTE]

• The primary ToolboxClient is asynchronous and requires using await for loading and invoking tools, as shown in most examples.
• Asynchronous code needs to run within an event loop (e.g., using asyncio.run() or in an async framework). See the Python asyncio documentation for more details.
• If you prefer synchronous execution, refer to the Synchronous Usage section below.

Quickstart

Here's a minimal example to get you started. Ensure your Toolbox service is running and accessible.

import asyncio
from toolbox_core import ToolboxClient

async def main():
    # Replace with the actual URL where your Toolbox service is running
    toolbox = ToolboxClient("http://127.0.0.1:5000")
    weather_tool = await toolbox.load_tool("get_weather")
    result = await weather_tool(location="London")
    print(result)

if __name__ == "__main__":
    asyncio.run(main())

Usage

Import and initialize a Toolbox client, pointing it to the URL of your running Toolbox service.

from toolbox_core import ToolboxClient

# Replace with your Toolbox service's URL
toolbox = ToolboxClient("http://127.0.0.1:5000")

All interactions for loading and invoking tools happen through this client.

Loading Tools

You can load tools individually or in groups (toolsets) as defined in your Toolbox service configuration. Loading a toolset is convenient when working with multiple related functions, while loading a single tool offers more granular control.

Load a toolset

A toolset is a collection of related tools. You can load all tools in a toolset or a specific one:

# Load all tools
tools = await toolbox.load_toolset()

# Load a specific toolset
tools = await toolbox.load_toolset("my-toolset")

Load a single tool

Loads a specific tool by its unique name. This provides fine-grained control.

tool = await toolbox.load_tool("my-tool")

Invoking Tools

Once loaded, tools behave like awaitable Python functions. You invoke them using await and pass arguments corresponding to the parameters defined in the tool's configuration within the Toolbox service.

tool = await toolbox.load_tool("my-tool")
result = await tool("foo", bar="baz")

[!TIP] For a more comprehensive guide on setting up the Toolbox service itself, which you'll need running to use this SDK, please refer to the Toolbox Quickstart Guide.

Synchronous Usage

By default, the ToolboxClient and the ToolboxTool objects it produces behave like asynchronous Python functions, requiring the use of await.

If your application primarily uses synchronous code, or you prefer not to manage an asyncio event loop, you can use the synchronous alternatives provided:

• ToolboxSyncClient: The synchronous counterpart to ToolboxClient.
• ToolboxSyncTool: The synchronous counterpart to ToolboxTool.

The ToolboxSyncClient handles communication with the Toolbox service synchronously and produces ToolboxSyncTool instances when you load tools. You do not use the await keyword when interacting with these synchronous versions.

from toolbox_core import ToolboxSyncClient

toolbox = ToolboxSyncClient("http://127.0.0.1:5000")
weather_tool = toolbox.load_tool("get_weather")
result = weather_tool(location="Paris")
print(result)

[!TIP] While synchronous invocation is available for convenience, it's generally considered best practice to use asynchronous operations (like those provided by the default ToolboxClient and ToolboxTool) for an I/O-bound task like tool invocation. Asynchronous programming allows for cooperative multitasking, often leading to better performance and resource utilization, especially in applications handling concurrent requests.

Use with LangGraph

The Toolbox Core SDK integrates smoothly with frameworks like LangGraph, allowing you to incorporate tools managed by the Toolbox service into your agentic workflows.

[!TIP] The loaded tools (both async ToolboxTool and sync ToolboxSyncTool) are callable and can often be used directly. However, to ensure parameter descriptions from Google-style docstrings are accurately parsed and made available to the LLM (via bind_tools()) and LangGraph internals, it's recommended to wrap the loaded tools using LangChain's StructuredTool.

Here's a conceptual example adapting the official LangGraph tool calling guide:

from toolbox_core import ToolboxClient
from langchain_google_vertexai import ChatVertexAI
from langgraph.graph import StateGraph, MessagesState, START, END
from langgraph.prebuilt import ToolNode
from langchain.tools import StructuredTool

toolbox = ToolboxClient("http://127.0.0.1:5000")
tools = await toolbox.load_toolset()
wrapped_tools = [StructuredTool.from_function(tool, parse_docstring=True) for tool in tools]
model_with_tools = ChatVertexAI(model="gemini-1.5-pro-002").bind_tools(wrapped_tools)

def call_model(state: MessagesState):
    messages = state["messages"]
    response = model_with_tools.invoke(messages)
    return {"messages": [response]}

def should_continue(state: MessagesState):
    messages = state["messages"]
    last_message = messages[-1]
    if last_message.tool_calls:
        return "tools"
    return END

workflow = StateGraph(MessagesState)

workflow.add_node("agent", call_model)
workflow.add_node("tools", ToolNode(wrapped_tools))

workflow.add_edge(START, "agent")
workflow.add_conditional_edges("agent", should_continue, ["tools", END])
workflow.add_edge("tools", "agent")

app = workflow.compile()

Client to Server Authentication

This section describes how to authenticate the ToolboxClient itself when connecting to a Toolbox server instance that requires authentication. This is crucial for securing your Toolbox server endpoint, especially when deployed on platforms like Cloud Run, GKE, or any environment where unauthenticated access is restricted.

This client-to-server authentication ensures that the Toolbox server can verify the identity of the client making the request before any tool is loaded or called. It is different from Authenticating Tools, which deals with providing credentials for specific tools within an already connected Toolbox session.

When is Client-to-Server Authentication Needed?

You'll need this type of authentication if your Toolbox server is configured to deny unauthenticated requests. For example:

• Your Toolbox server is deployed on Cloud Run and configured to "Require authentication."
• Your server is behind an Identity-Aware Proxy (IAP) or a similar authentication layer.
• You have custom authentication middleware on your self-hosted Toolbox server.

Without proper client authentication in these scenarios, attempts to connect or make calls (like load_tool) will likely fail with Unauthorized errors.

How it works

The ToolboxClient (and ToolboxSyncClient) allows you to specify functions (or coroutines for the async client) that dynamically generate HTTP headers for every request sent to the Toolbox server. The most common use case is to add an Authorization header with a bearer token (e.g., a Google ID token).

These header-generating functions are called just before each request, ensuring that fresh credentials or header values can be used.

Configuration

You can configure these dynamic headers in two ways:

1. During Client Initialization

   from toolbox_core import ToolboxClient
   
   client = ToolboxClient("toolbox-url", headers={"header1": header1_getter, "header2": header2_getter, ...})
   

2. After Client Initialization

   from toolbox_core import ToolboxClient
   
   client = ToolboxClient("toolbox-url")
   client.add_headers({"header1": header1_getter, "header2": header2_getter, ...})
   

Authenticating with Google Cloud Servers

For Toolbox servers hosted on Google Cloud (e.g., Cloud Run) and requiring Google ID token authentication, the helper module auth_methods provides utility functions.

Step by Step Guide for Cloud Run

1. Configure Permissions: Grant the roles/run.invoker IAM role on the Cloud Run service to the principal. This could be your user account email or a service account.

2. Configure Credentials

   • Local Development: Set up ADC.
   • Google Cloud Environments: When running within Google Cloud (e.g., Compute Engine, GKE, another Cloud Run service, Cloud Functions), ADC is typically configured automatically, using the environment's default service account.

3. Connect to the Toolbox Server

   from toolbox_core import auth_methods
   
   auth_token_provider = auth_methods.aget_google_id_token # can also use sync method
   client = ToolboxClient(
       URL,
       client_headers={"Authorization": auth_token_provider},
   )
   tools = await client.load_toolset()
   
   # Now, you can use the client as usual.
   

Authenticating Tools

[!WARNING] Always use HTTPS to connect your application with the Toolbox service, especially in production environments or whenever the communication involves sensitive data (including scenarios where tools require authentication tokens). Using plain HTTP lacks encryption and exposes your application and data to significant security risks, such as eavesdropping and tampering.

Tools can be configured within the Toolbox service to require authentication, ensuring only authorized users or applications can invoke them, especially when accessing sensitive data.

When is Authentication Needed?

Authentication is configured per-tool within the Toolbox service itself. If a tool you intend to use is marked as requiring authentication in the service, you must configure the SDK client to provide the necessary credentials (currently Oauth2 tokens) when invoking that specific tool.

Supported Authentication Mechanisms

The Toolbox service enables secure tool usage through Authenticated Parameters. For detailed information on how these mechanisms work within the Toolbox service and how to configure them, please refer to Toolbox Service Documentation - Authenticated Parameters

Step 1: Configure Tools in Toolbox Service

First, ensure the target tool(s) are configured correctly in the Toolbox service to require authentication. Refer to the Toolbox Service Documentation - Authenticated Parameters for instructions.

Step 2: Configure SDK Client

Your application needs a way to obtain the required Oauth2 token for the authenticated user. The SDK requires you to provide a function capable of retrieving this token when the tool is invoked.

Provide an ID Token Retriever Function

You must provide the SDK with a function (sync or async) that returns the necessary token when called. The implementation depends on your application's authentication flow (e.g., retrieving a stored token, initiating an OAuth flow).

[!IMPORTANT] The name used when registering the getter function with the SDK (e.g., "my_api_token") must exactly match the name of the corresponding authServices defined in the tool's configuration within the Toolbox service.

async def get_auth_token():
    # ... Logic to retrieve ID token (e.g., from local storage, OAuth flow)
    # This example just returns a placeholder. Replace with your actual token retrieval.
    return "YOUR_ID_TOKEN" # Placeholder

[!TIP] Your token retriever function is invoked every time an authenticated parameter requires a token for a tool call. Consider implementing caching logic within this function to avoid redundant token fetching or generation, especially for tokens with longer validity periods or if the retrieval process is resource-intensive.

Option A: Add Authentication to a Loaded Tool

You can add the token retriever function to a tool object after it has been loaded. This modifies the specific tool instance.

toolbox = ToolboxClient("http://127.0.0.1:5000")
tool = await toolbox.load_tool("my-tool")

auth_tool = tool.add_auth_token_getter("my_auth", get_auth_token)  # Single token

# OR

multi_auth_tool = tool.add_auth_token_getters({
    "my_auth_1", get_auth_token_1,
    "my_auth_2", get_auth_token_2,
})  # Multiple tokens

Option B: Add Authentication While Loading Tools

You can provide the token retriever(s) directly during the load_tool or load_toolset calls. This applies the authentication configuration only to the tools loaded in that specific call, without modifying the original tool objects if they were loaded previously.

auth_tool = await toolbox.load_tool(auth_token_getters={"my_auth": get_auth_token})

# OR

auth_tools = await toolbox.load_toolset(auth_token_getters={"my_auth": get_auth_token})

[!NOTE] Adding auth tokens during loading only affect the tools loaded within that call.

Complete Authentication Example

import asyncio
from toolbox_core import ToolboxClient

async def get_auth_token():
    # ... Logic to retrieve ID token (e.g., from local storage, OAuth flow)
    # This example just returns a placeholder. Replace with your actual token retrieval.
    return "YOUR_ID_TOKEN" # Placeholder

toolbox = ToolboxClient("http://127.0.0.1:5000")
tool = await toolbox.load_tool("my-tool")

auth_tool = tool.add_auth_token_getters({"my_auth": get_auth_token})
result = auth_tool(input="some input")
print(result)

Binding Parameter Values

The SDK allows you to pre-set, or "bind", values for specific tool parameters before the tool is invoked or even passed to an LLM. These bound values are fixed and will not be requested or modified by the LLM during tool use.

Why Bind Parameters?

• Protecting sensitive information: API keys, secrets, etc.
• Enforcing consistency: Ensuring specific values for certain parameters.
• Pre-filling known data: Providing defaults or context.

[!IMPORTANT] The parameter names used for binding (e.g., "api_key") must exactly match the parameter names defined in the tool's configuration within the Toolbox service.

[!NOTE] You do not need to modify the tool's configuration in the Toolbox service to bind parameter values using the SDK.

Option A: Binding Parameters to a Loaded Tool

Bind values to a tool object after it has been loaded. This modifies the specific tool instance.

toolbox = ToolboxClient("http://127.0.0.1:5000")
tool = await toolbox.load_tool("my-tool")

bound_tool = tool.bind_param("param", "value")

# OR

bound_tool = tool.bind_params({"param": "value"})

Option B: Binding Parameters While Loading Tools

Specify bound parameters directly when loading tools. This applies the binding only to the tools loaded in that specific call.

bound_tool = await toolbox.load_tool("my-tool", bound_params={"param": "value"})

# OR

bound_tools = await toolbox.load_toolset(bound_params={"param": "value"})

[!NOTE] Bound values during loading only affect the tools loaded in that call.

Binding Dynamic Values

Instead of a static value, you can bind a parameter to a synchronous or asynchronous function. This function will be called each time the tool is invoked to dynamically determine the parameter's value at runtime.

async def get_dynamic_value():
  # Logic to determine the value
  return "dynamic_value"

dynamic_bound_tool = tool.bind_param("param", get_dynamic_value)

[!IMPORTANT] You don't need to modify tool configurations to bind parameter values.

Contributing

Contributions are welcome! Please refer to the DEVELOPER.md file for guidelines on how to set up a development environment and run tests.

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Support

If you encounter issues or have questions, check the existing GitHub Issues for the main Toolbox project.