Native integration with AWS Bedrock LLMs

[br-sathwik]

Feature Request: Native integration with AWS Bedrock LLMs

Required Information

Is your feature request related to a specific problem?

ADK's native model support is currently limited to Google Gemini and a small number of third-party providers. This prevents AWS Bedrock users from seamlessly integrating Bedrock models, including Amazon Nova, Claude, Llama, and Mistral. While LiteLLM-based integration is available, it does not support bidirectional streaming for Speech-to-Speech (STS) models such as Nova Sonic. A native Bedrock integration would eliminate the limitations of intermediary layers, provide first-class support for Bedrock models and features, and ensure long-term compatibility as both ADK and AWS AI services evolve.

Describe the Solution You'd Like

Add a first-class BedrockLlm provider that implements BaseLlm and BaseLlmConnection, supporting:

• Text generation (non-streaming) via Bedrock's converse() API
• Text generation (streaming) via Bedrock's converse_stream() API
• Bidirectional streaming (voice) via Bedrock's invoke_model_with_bidirectional_stream() API
• Tool/function calling with full schema support
• Prompt caching for system prompts and tool definitions
• All Bedrock Converse-compatible models: Amazon Nova, Anthropic Claude, Meta Llama, Mistral, Cohere, AI21 Labs, and cross-region variants

The provider would use the aws-sdk-bedrock-runtime Python SDK and standard AWS credential configuration (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION).

Usage example:

from google.adk.models.bedrock import BedrockLlm

# Text model
llm = BedrockLlm(model="amazon.nova-pro-v1:0")

# Voice model (bidirectional streaming)
llm = BedrockLlm(model="amazon.nova-2-sonic-v1:0")

Impact on your work

This enables ADK adoption for developers and teams using AWS. Without native Bedrock support, users cannot leverage ADK with Bedrock models or must maintain custom integrations that break across ADK version upgrades.

Willingness to contribute

Yes. We have a working implementation with full test coverage that is ready to be submitted as a PR.

---

Recommended Information

Describe Alternatives You've Considered

1. Custom BaseLlm subclass per project - Duplicates effort across users and is hard to keep in sync with ADK changes.
2. LiteLLM - Adds an extra hop, doesn't support bidirectional streaming for voice models, and introduces cold-start latency.

Proposed API / Implementation

from google.adk.models.bedrock import BedrockLlm, BedrockLlmConnection

# BedrockLlm implements BaseLlm
# - generate_content_async() -> uses converse() or converse_stream()
# - connect() -> returns BedrockLlmConnection for bidi streaming
# - supported_models() -> auto-detects Amazon Nova model IDs

# BedrockLlmConnection implements BaseLlmConnection
# - Manages HTTP/2 full-duplex stream lifecycle
# - Parses Bedrock events (audioOutput, textOutput, toolUse, etc.)
# - Handles interruptions and stream cancellation

Key implementation details:

• Singleton BedrockRuntimeClient shared across instances
• Tool schema normalization (type lowercasing, Decimal-to-float coercion)

Additional Context

• Tested with Python 3.12 and 3.13
• Licensed under Apache-2.0 (same as ADK)
• All models available via Bedrock Converse API are supported by passing the model ID directly