Amazon Bedrock and Deepgram Voice Agent

This guide walks you through setting up a proxy server to route Deepgram Voice Agent’s LLM requests to Amazon Bedrock, enabling you to use Bedrock’s language models with Deepgram’s Voice agent.

Before you Begin

1. Prerequisites

You will need:

• An understanding of Python and using Python virtual environments.
• An Amazon Bedrock service account
• A Deepgram Voice Agent. Here’s our guide on building the voice agent.
• ngrok to allow access to a local server OR your own hosted server

2. Architecture Overview

How it works:

• The proxy logs and forwards agent.think payloads to Bedrock
• Bedrock handles LLM logic and returns structured responses
• Deepgram converts the response into speech back to the user

3. Set Up the Proxy

Clone the proxy repo

$
git clone https://github.com/deepgram-devs/deepgram-voice-agent-client-llm-proxy.git
>
cd deepgram-voice-agent-client-llm-proxy
>
python3 -m venv venv
>
source venv/bin/activate
>
pip install -r requirements.txt

Configure the environment

$
cp .env.example .env

Specifying Bedrock provider details

1
AGENT_ID=your_bedrock_agent_id
2
AGENT_ALIAS_ID=your_bedrock_agent_alias_id
3
AWS_ACCESS_KEY_ID=your_aws_access_key_id
4
AWS_SECRET_ACCESS_KEY=your_aws_secret_access_key
5
AWS_REGION=us-east-1

Start the server

$
python app.py

Using ngrok

Be sure to set the port correctly to 5000 by running:

$
ngrok http 5000

4. Configure Deepgram Voice Agent

In your Deepgram Voice Agent settings, update the provider, model, and endpoint URL for agent.think. See more examples of configuration here.

1
"agent": {
2
  "think": {
3
    "provider": {
4
      "type": "open_ai",
5
      "model": "gpt-4o-mini",
6
      "temperature": 0.7
7
    },
8
    "endpoint": {
9
      "url": "{{host}}/v1/chat/completions",
10
      "headers": {
11
        "authorization": "Bearer {{token}}"
12
      }
13
    }
14
  }
15
}

5. Test the Integration

1. Launch the proxy and ngrok
2. Deploy your Deepgram Voice Agent with the updated config
3. Start a call or session
4. Observe agent.think payloads and Bedrock responses in proxy logs
5. Confirm LLM responses originate from Bedrock (e.g., function calls reflected)