Thanks — I see you want the realtime LLM to always perform a tool call instead of sometimes replying directly. Yes — set the model's

tool_choice

parameter to

"required"

. That forces the LLM to make a tool call every turn and prevents it from returning a normal message. Example (Python-style):

llm = LLM(
    model="gpt-4o",
    tool_choice="required"
)

Or, when creating an Agent session:

session = AgentSession(
    llm=openai.LLM.with_deepseek(
        model="deepseek-chat",
        tool_choice="required",
    ),
    # ...
)

If you want, tell me which model/plugin (OpenAI Realtime, DeepSeek, Groq, Anthropic, etc.) and how you’re configuring the agent (code snippet or config), and I’ll give exact code for that plugin. Sources: DeepSeek LLM plugin guide | LiveKit docs | Groq LLM plugin guide | LiveKit docs | Anthropic Claude LLM plugin guide | LiveKit docs | xAI LLM plugin guide | LiveKit docs | Azure OpenAI LLM plugin guide | LiveKit docs