Skip to main content

Documentation Index

Fetch the complete documentation index at: https://doc.rapida.ai/llms.txt

Use this file to discover all available pages before exploring further.

AgentKit lets you run your own LLM logic on your own server. Instead of using Rapida’s built-in LLM endpoint routing, Rapida calls your gRPC service in real time — streaming user speech transcripts to your server, which streams back assistant responses that are synthesized to audio and played to the caller. This gives you complete control over:
  • Which LLM you call and how you call it
  • Custom tool definitions and execution
  • Multi-step reasoning chains (LangChain, CrewAI, AutoGen, etc.)
  • State management and memory
  • Any external API or database your logic needs

​
How It Works

User speaks
    ↓ (Rapida transcribes via STT)
Your AgentKit server receives TalkInput (user text)
    ↓ (your code calls OpenAI / Anthropic / custom model)
Your server streams back TalkOutput (assistant text chunks)
    ↓ (Rapida synthesizes via TTS and plays to user)
User hears the response
The protocol is a bidirectional gRPC stream. Rapida manages the audio pipeline (VAD, STT, TTS, telephony) — your server only handles text in / text out.

​
Python Quick Start

​
1. Install the SDK

pip install rapida-python

​
2. Implement your agent

Subclass AgentKitAgent and override the Talk method: 
import os
from openai import OpenAI
from rapida import AgentKitAgent, AgentKitServer
from rapida.clients.protos.talk_api_pb2 import TalkInput, TalkOutput

openai_client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

class MyVoiceAgent(AgentKitAgent):
    def Talk(self, request_iterator, context):
        messages = [{"role": "system", "content": "You are a helpful voice assistant."}]

        for request in request_iterator:
            # Acknowledge configuration handshake
            if self.is_configuration_request(request):
                yield self.configuration_response(request.configuration)
                continue

            # Get the user's transcribed text
            user_text = self.get_user_text(request)
            if not user_text:
                continue

            messages.append({"role": "user", "content": user_text})
            msg_id = self.get_message_id(request)

            # Stream the LLM response back
            full_response = ""
            stream = openai_client.chat.completions.create(
                model="gpt-4o",
                messages=messages,
                stream=True,
            )
            for chunk in stream:
                delta = chunk.choices[0].delta.content or ""
                if delta:
                    full_response += delta
                    yield self.assistant_response(msg_id, delta, completed=False)

            # Signal end of response
            yield self.assistant_response(msg_id, "", completed=True)
            messages.append({"role": "assistant", "content": full_response})


if __name__ == "__main__":
    server = AgentKitServer(
        agent=MyVoiceAgent(),
        host="0.0.0.0",
        port=50051,
    )
    print("AgentKit server listening on port 50051")
    server.start()
    server.wait_for_termination()

​
3. Point your assistant to your server

In the Rapida dashboard, when configuring your assistant’s LLM provider, select AgentKit and enter the address of your server (e.g. my-server.example.com:50051).

​
Response Types

Your Talk method yields TalkOutput objects. Use the convenience methods on AgentKitAgent:
MethodUse when
assistant_response(msg_id, text, completed)Streaming text response chunks
configuration_response(config)Acknowledging the initial handshake
tool_call(msg_id, tool_id, name, args)Requesting Rapida to execute a tool
tool_call_result(msg_id, tool_id, name, result, success)Returning a tool result
transfer_call(msg_id, args)Transferring the call to another number
terminate_call(msg_id, args)Ending the call programmatically
error_response(code, message)Signalling an error

​
Handling Tool Calls

Rapida can execute tools on your behalf (knowledge retrieval, endpoint invocation) and send results back to your server. Here’s how to request a tool call and receive its result: 
for request in request_iterator:
    if self.is_configuration_request(request):
        yield self.configuration_response(request.configuration)
        continue

    user_text = self.get_user_text(request)
    msg_id = self.get_message_id(request)

    # Request a tool call from Rapida
    yield self.tool_call(
        msg_id=msg_id,
        tool_id="tool_001",
        name="get_weather",
        args={"location": "London"}
    )

    # The next message will contain the tool result
    tool_result_request = next(request_iterator)
    # Process tool result and continue...

​
SSL / TLS Configuration

For production deployments, enable TLS: 
server = AgentKitServer(
    agent=MyVoiceAgent(),
    host="0.0.0.0",
    port=50051,
    ssl_config={
        "cert_path": "/etc/ssl/server.crt",
        "key_path": "/etc/ssl/server.key",
        "ca_cert_path": "/etc/ssl/ca.crt",  # Optional: for mutual TLS
    }
)

​
Authentication

Protect your server with a bearer token: 
server = AgentKitServer(
    agent=MyVoiceAgent(),
    host="0.0.0.0",
    port=50051,
    auth_config={
        "enabled": True,
        "token": os.getenv("AGENTKIT_TOKEN"),
    }
)

​
Framework Examples

AgentKit works with any Python code. Here are some common patterns: 
from langchain.agents import AgentExecutor, create_react_agent
from langchain_openai import ChatOpenAI
from langchain.tools import tool

@tool
def get_account_balance(account_id: str) -> str:
    """Get the balance for a customer account."""
    # Call your database or API
    return f"Account {account_id} balance: $1,234.56"

llm = ChatOpenAI(model="gpt-4o", streaming=True)
tools = [get_account_balance]
agent = create_react_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools)

class LangChainAgent(AgentKitAgent):
    def Talk(self, request_iterator, context):
        for request in request_iterator:
            if self.is_configuration_request(request):
                yield self.configuration_response(request.configuration)
                continue
            user_text = self.get_user_text(request)
            msg_id = self.get_message_id(request)
            for chunk in agent_executor.stream({"input": user_text}):
                if "output" in chunk:
                    yield self.assistant_response(msg_id, chunk["output"], completed=False)
            yield self.assistant_response(msg_id, "", completed=True)

import anthropic
client = anthropic.Anthropic()

class ClaudeAgent(AgentKitAgent):
    def Talk(self, request_iterator, context):
        messages = []
        for request in request_iterator:
            if self.is_configuration_request(request):
                yield self.configuration_response(request.configuration)
                continue
            user_text = self.get_user_text(request)
            msg_id = self.get_message_id(request)
            messages.append({"role": "user", "content": user_text})
            with client.messages.stream(
                model="claude-opus-4-6",
                max_tokens=1024,
                messages=messages,
            ) as stream:
                full = ""
                for text in stream.text_stream:
                    full += text
                    yield self.assistant_response(msg_id, text, completed=False)
            yield self.assistant_response(msg_id, "", completed=True)
            messages.append({"role": "assistant", "content": full})