Connect Arcade to your LLM

Outcomes

Integrate Arcade -calling capabilities into an application that uses an LLM in Python.

You will Learn

Setup a simple agentic loop
Add Arcade to your agentic loop
Implement a multi-turn conversation loop

Prerequisites

Create a new project and install the dependencies

In your terminal, run the following command to create a new uv

Terminal


mkdir arcade-llm-example
cd arcade-llm-example
uv init

Create a new virtual environment and activate it:

Terminal


uv venv
source .venv/bin/activate

Install the dependencies:

Terminal


uv add arcadepy openai python-dotenv

Your directory should now look like this:

Terminal


arcade-llm-example/
├── .git/
├── .gitignore
├── python-version
├── .venv/
├── main.py
├── pyproject.toml
├── main.py
├── README.md
└── uv.lock

Instantiate and use the clients

Create a new file called .env and add your Arcade , as well as your OpenAI API key:

Terminal


ARCADE_API_KEY=YOUR_ARCADE_API_KEY
ARCADE_USER_ID=YOUR_ARCADE_USER_ID
OPENROUTER_API_KEY=YOUR_OPENROUTER_API_KEY
OPENROUTER_MODEL=YOUR_OPENROUTER_MODEL

In this example, we’re using OpenRouter to access the model, as it makes it very easy to use any model from multiple providers with a single API. OpenRouter is compliant with the OpenAI API specification, so you can use it with any OpenAI-compatible library.

Open the main.py file in your editor of choice, and replace replace the contents with the following:

Python

main.py


from arcadepy import Arcade
from openai import OpenAI
from dotenv import load_dotenv
import json
import os
 
load_dotenv()
 
client = Arcade()
arcade_user_id = os.getenv("ARCADE_USER_ID")
llm_client = OpenAI(
    api_key=os.getenv("OPENROUTER_API_KEY"),
    base_url="https://openrouter.ai/api/v1"
)

Select and retrieve the tools from Arcade

In this example, we’re implementing a simple that can retrieve and send emails, as well as send messages to Slack. In a harness like this one, where the entire catalog of is exposed to the LLM, you will want to choose only the tools that are relevant to the task at hand to avoid overwhelming the LLM with too many options, and to make your agent more token efficient.

Python

main.py


# We define here the tools that we want to use in the agent
tool_catalog = [
    "Gmail.ListEmails",
    "Gmail.SendEmail",
    "Slack.SendMessage",
    "Slack.WhoAmI"
]
 
# We get the tool definitions from the Arcade API, so that we can expose them to the LLM
tool_definitions = []
for tool in tool_catalog:
    tool_definitions.append(arcade.tools.formatted.get(name=tool, format="openai"))

Write a helper function that handles tool authorization and execution

The LLM may choose to call any of the tools in our catalog, and we need to handle different scenarios, such as when the tool requires authorization. There are many approaches to this, but a good pattern is to handle interruptions like this outside of the agentic if we can do so. As a rule of thumb, you should evaluate whether it is relevant for the LLM to be aware of the authorization process, or if it’s better to handle it in the harness, and keep the context as if the was already authorized. The latter option optimizes for token efficiency, and we will implement it in this example.

Python

main.py


# Helper function to authorize and run any tool
def authorize_and_run_tool(tool_name: str, input: str):
    # Start the authorization process
    auth_response = arcade.tools.authorize(
        tool_name=tool_name,
        user_id=arcade_user_id,
    )
 
    # If the authorization is not completed, print the authorization URL and wait for the user to authorize the app.
    # Tools that do not require authorization will have the status "completed" already.
    if auth_response.status != "completed":
        print(f"Click this link to authorize {tool_name}: {auth_response.url}. The process will continue once you have authorized the app.")
        arcade.auth.wait_for_completion(auth_response.id)
 
    # Parse the input
    input_json = json.loads(input)
 
    # Run the tool
    result = arcade.tools.execute(
        tool_name=tool_name,
        input=input_json,
        user_id=arcade_user_id,
    )
 
    # Return the tool output to the caller as a JSON string
    return json.dumps(result.output.value)

That helper function adapts to any tool in the catalog, and will make sure that the authorization requirements are met before executing the tool. For more complex agentic patterns, this is generally the best place to handle interruptions that may require user interaction, such as when the tool requires a user to approve a request, or to provide additional .

Write a helper function that handles the LLM’s invocation

There are many orchestration patterns that can be used to handle the LLM invocation. A common pattern is a ReAct architecture, where the user prompt will result in a loop of messages between the LLM and the tools, until the LLM provides a final response (no calls). This is the pattern we will implement in this example.

To avoid the risk of infinite loops, we will limit the number of turns to a maximum of 5. This is a parameter that you can tune to your needs, and it’s a good idea to set it to a value that is high enough to allow the LLM to complete the task it’s designed to do, but low enough to prevent infinite loops.

Python

main.py


def invoke_llm(
    history: list[dict],
    model: str = "google/gemini-2.5-flash",
    max_turns: int = 5,
    tools: list[dict] = None,
    tool_choice: str = "auto",
) -> list[dict]:
    """
    Multi-turn LLM invocation that processes the conversation until
    the assistant provides a final response (no tool calls).
 
    Returns the updated conversation history.
    """
    turns = 0
 
    while turns < max_turns:
        turns += 1
 
        response = llm_client.chat.completions.create(
            model=model,
            messages=history,
            tools=tools,
            tool_choice=tool_choice,
        )
 
        assistant_message = response.choices[0].message
 
        if assistant_message.tool_calls:
            for tool_call in assistant_message.tool_calls:
                tool_name = tool_call.function.name
                tool_args = tool_call.function.arguments
                print(f"🛠️ Harness: Calling {tool_name} with input {tool_args}")
                tool_result = authorize_and_run_tool(tool_name, tool_args)
                print(f"🛠️ Harness: Tool call {tool_name} completed")
                history.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "content": tool_result,
                })
 
            continue
 
        else:
            history.append({
                "role": "assistant",
                "content": assistant_message.content,
            })
 
        break
 
    return history

In combination, these two helper functions will form the core of our agentic loop. You will notice that the authorization is handled outside of the agentic context, and the execution is passed back to the LLM in every case. Depending on your needs, you may want to handle tool orchestration within the harness, and pass only the final result of multiple tool calls to the LLM.

Write the main agentic loop

Now that we’ve written the helper functions, we can write a very simple agentic loop that interacts with the . The core pieces of this loop are:

Initialize the conversation history with the system prompt
Get the input, and add it to the conversation history
Invoke the LLM with the conversation history, tools, and choice
Repeat from step 2 until the decides to stop the conversation

Python

main.py


def chat():
    """Interactive multi-turn chat session."""
    print("Chat started. Type 'quit' or 'exit' to end the session.\n")
 
    # Initialize the conversation history with the system prompt
    history: list[dict] = [
        {"role": "system", "content": "You are a helpful assistant."}
    ]
 
    while True:
        try:
            user_input = input("😎 You: ").strip()
        except (EOFError, KeyboardInterrupt):
            print("\nGoodbye!")
            break
 
        if not user_input:
            continue
 
        if user_input.lower() in ("quit", "exit"):
            print("Goodbye!")
            break
 
        # Add user message to history
        history.append({"role": "user", "content": user_input})
 
        # Get LLM response
        history = invoke_llm(
            history, tools=tool_definitions)
 
        # Print the latest assistant response
        assistant_response = history[-1]["content"]
        print(f"\n🤖 Assistant: {assistant_response}\n")
 
 
if __name__ == "__main__":
    chat()

Run the code

It’s time to run the code and see it in action! Run the following command to start the chat:

Terminal


uv run main.py

With the selection of tools above, you should be able to get the to effectively complete the following prompts:

“Please send a message to the #general channel on Slack greeting everyone with a haiku about .”
“Please write a poem about multi- orchestration and send it to the #general channel on Slack, also send it to me in an email.”
“Please summarize my latest 5 emails, then send me a DM on Slack with the summary.”

Next Steps

Learn more about using Arcade with frameworks like LangChain or Mastra.
Learn more about how to build your own MCP Servers.

Example code

Python

main.py


from arcadepy import Arcade
from dotenv import load_dotenv
from openai import OpenAI
import json
import os
 
load_dotenv()
 
arcade = Arcade()
arcade_user_id = os.getenv("ARCADE_USER_ID")
llm_client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.getenv("OPENROUTER_API_KEY"),
)
 
# We define here the tools that we want to use in the agent
tool_catalog = [
    "Gmail.ListEmails",
    "Gmail.SendEmail",
    "Slack.SendMessage",
    "Slack.WhoAmI"
]
 
# We get the tool definitions from the Arcade API, so that we can expose them to the LLM
tool_definitions = []
for tool in tool_catalog:
    tool_definitions.append(arcade.tools.formatted.get(name=tool, format="openai"))
 
 
# Helper function to authorize and run any tool
def authorize_and_run_tool(tool_name: str, input: str):
    # Start the authorization process
    auth_response = arcade.tools.authorize(
        tool_name=tool_name,
        user_id=arcade_user_id,
    )
 
    # If the authorization is not completed, print the authorization URL and wait for the user to authorize the app.
    # Tools that do not require authorization will have the status "completed" already.
    if auth_response.status != "completed":
        print(f"Click this link to authorize {tool_name}: {auth_response.url}. The process will continue once you have authorized the app.")
        arcade.auth.wait_for_completion(auth_response.id)
 
    # Parse the input
    input_json = json.loads(input)
 
    # Run the tool
    result = arcade.tools.execute(
        tool_name=tool_name,
        input=input_json,
        user_id=arcade_user_id,
    )
 
    # Return the tool output to the caller as a JSON string
    return json.dumps(result.output.value)
 
 
def invoke_llm(
    history: list[dict],
    model: str = "google/gemini-2.5-flash",
    max_turns: int = 5,
    tools: list[dict] = None,
    tool_choice: str = "auto",
) -> list[dict]:
    """
    Multi-turn LLM invocation that processes the conversation until
    the assistant provides a final response (no tool calls).
 
    Returns the updated conversation history.
    """
    turns = 0
 
    while turns < max_turns:
        turns += 1
 
        response = llm_client.chat.completions.create(
            model=model,
            messages=history,
            tools=tools,
            tool_choice=tool_choice,
        )
 
        assistant_message = response.choices[0].message
 
        if assistant_message.tool_calls:
            for tool_call in assistant_message.tool_calls:
                tool_name = tool_call.function.name
                tool_args = tool_call.function.arguments
                print(f"🛠️ Harness: Calling {tool_name} with input {tool_args}")
                tool_result = authorize_and_run_tool(tool_name, tool_args)
                print(f"🛠️ Harness: Tool call {tool_name} completed")
                history.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "content": tool_result,
                })
 
            continue
 
        else:
            history.append({
                "role": "assistant",
                "content": assistant_message.content,
            })
 
        break
 
    return history
 
 
def chat():
    """Interactive multi-turn chat session."""
    print("Chat started. Type 'quit' or 'exit' to end the session.\n")
 
    history: list[dict] = [
        {"role": "system", "content": "You are a helpful assistant."}
    ]
 
    while True:
        try:
            user_input = input("😎 You: ").strip()
        except (EOFError, KeyboardInterrupt):
            print("\nGoodbye!")
            break
 
        if not user_input:
            continue
 
        if user_input.lower() in ("quit", "exit"):
            print("Goodbye!")
            break
 
        # Add user message to history
        history.append({"role": "user", "content": user_input})
 
        # Get LLM response
        history = invoke_llm(
            history, tools=tool_definitions)
 
        # Print the latest assistant response
        assistant_response = history[-1]["content"]
        print(f"\n🤖 Assistant: {assistant_response}\n")
 
 
if __name__ == "__main__":
    chat()