[Feat] LiteLLM x Claude Agent SDK Integration (#20035)

* fix: bedrock invoke - does not support prompt-caching-scope * fix: UNSUPPORTED_BEDROCK_INVOKE_BETA_PATTERNS * init requirements.txt * init README for claude Agent SDK * fix: using converse models with UNSUPPORTED_BEDROCK_CONVERSE_BETA_PATTERNS * fix main.py * init: proxy_e2e_anthropic_messages_tests
2026-08-02 06:22:48 +00:00 · 2026-01-29 17:48:38 -08:00
parent f7e1a22947
commit 476f0b29d2
10 changed files with 669 additions and 96 deletions
@@ -3407,6 +3407,110 @@ jobs:
      - store_test_results:
          path: test-results

+  proxy_e2e_anthropic_messages_tests:
+    machine:
+      image: ubuntu-2204:2023.10.1
+    resource_class: xlarge
+    working_directory: ~/project
+    steps:
+      - checkout
+      - setup_google_dns
+      - run:
+          name: Install Docker CLI (In case it's not already installed)
+          command: |
+            curl -fsSL https://get.docker.com | sh
+            sudo usermod -aG docker $USER
+            docker version
+      - run:
+          name: Install Python 3.10
+          command: |
+            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh --output miniconda.sh
+            bash miniconda.sh -b -p $HOME/miniconda
+            export PATH="$HOME/miniconda/bin:$PATH"
+            conda init bash
+            source ~/.bashrc
+            conda create -n myenv python=3.10 -y
+            conda activate myenv
+            python --version
+      - run:
+          name: Install Dependencies
+          command: |
+            export PATH="$HOME/miniconda/bin:$PATH"
+            source $HOME/miniconda/etc/profile.d/conda.sh
+            conda activate myenv
+            pip install "pytest==7.3.1"
+            pip install "pytest-asyncio==0.21.1"
+            pip install "boto3==1.36.0"
+            pip install "httpx==0.27.0"
+            pip install "claude-agent-sdk"
+            pip install -r requirements.txt
+      - run:
+          name: Install dockerize
+          command: |
+            wget https://github.com/jwilder/dockerize/releases/download/v0.6.1/dockerize-linux-amd64-v0.6.1.tar.gz
+            sudo tar -C /usr/local/bin -xzvf dockerize-linux-amd64-v0.6.1.tar.gz
+            rm dockerize-linux-amd64-v0.6.1.tar.gz
+      - run:
+          name: Start PostgreSQL Database
+          command: |
+            docker run -d \
+              --name postgres-db \
+              -e POSTGRES_USER=postgres \
+              -e POSTGRES_PASSWORD=postgres \
+              -e POSTGRES_DB=circle_test \
+              -p 5432:5432 \
+              postgres:14
+      - run:
+          name: Wait for PostgreSQL to be ready
+          command: dockerize -wait tcp://localhost:5432 -timeout 1m
+      - attach_workspace:
+          at: ~/project
+      - run:
+          name: Load Docker Database Image
+          command: |
+            gunzip -c litellm-docker-database.tar.gz | docker load
+            docker images | grep litellm-docker-database
+      - run:
+          name: Run Docker container with test config
+          command: |
+            docker run -d \
+              -p 4000:4000 \
+              -e DATABASE_URL=postgresql://postgres:postgres@host.docker.internal:5432/circle_test \
+              -e LITELLM_MASTER_KEY="sk-1234" \
+              -e AWS_ACCESS_KEY_ID=$AWS_ACCESS_KEY_ID \
+              -e AWS_SECRET_ACCESS_KEY=$AWS_SECRET_ACCESS_KEY \
+              -e AWS_REGION_NAME="us-east-1" \
+              --add-host host.docker.internal:host-gateway \
+              --name my-app \
+              -v $(pwd)/tests/proxy_e2e_anthropic_messages_tests/test_config.yaml:/app/config.yaml \
+              litellm-docker-database:ci \
+              --config /app/config.yaml \
+              --port 4000 \
+              --detailed_debug
+      - run:
+          name: Start outputting logs
+          command: docker logs -f my-app
+          background: true
+      - run:
+          name: Wait for app to be ready
+          command: dockerize -wait http://localhost:4000 -timeout 5m
+      - run:
+          name: Run Claude Agent SDK E2E Tests
+          command: |
+            export PATH="$HOME/miniconda/bin:$PATH"
+            source $HOME/miniconda/etc/profile.d/conda.sh
+            conda activate myenv
+            export LITELLM_PROXY_URL="http://localhost:4000"
+            export LITELLM_API_KEY="sk-1234"
+            pwd
+            ls
+            python -m pytest -vv tests/proxy_e2e_anthropic_messages_tests/ -x -s --junitxml=test-results/junit.xml --durations=5
+          no_output_timeout: 120m
+
+      # Store test results
+      - store_test_results:
+          path: test-results
+
  upload-coverage:
    docker:
      - image: cimg/python:3.9
@@ -4075,6 +4179,14 @@ workflows:
              only:
                - main
                - /litellm_.*/
+      - proxy_e2e_anthropic_messages_tests:
+          requires:
+            - build_docker_database_image
+          filters:
+            branches:
+              only:
+                - main
+                - /litellm_.*/
      - llm_translation_testing:
          filters:
            branches:
@@ -0,0 +1,117 @@
+# Claude Agent SDK with LiteLLM Gateway
+
+A simple example showing how to use Claude's Agent SDK with LiteLLM as a proxy. This lets you use any LLM provider (OpenAI, Bedrock, Azure, etc.) through the Agent SDK.
+
+## Quick Start
+
+### 1. Install dependencies
+
+```bash
+pip install anthropic claude-agent-sdk litellm
+```
+
+### 2. Start LiteLLM proxy
+
+```bash
+# Simple start with Claude
+litellm --model claude-sonnet-4-20250514
+
+# Or with a config file
+litellm --config config.yaml
+```
+
+### 3. Run the chat
+
+```bash
+python main.py
+```
+
+That's it! You can now chat with the agent in your terminal.
+
+### Chat Commands
+
+While chatting, you can use these commands:
+- `models` - List all available models (fetched from your LiteLLM proxy)
+- `model` - Switch to a different model
+- `clear` - Start a new conversation
+- `quit` or `exit` - End the chat
+
+The chat automatically fetches available models from your LiteLLM proxy's `/models` endpoint, so you'll always see what's currently configured.
+
+## Configuration
+
+Set these environment variables if needed:
+
+```bash
+export LITELLM_PROXY_URL="http://localhost:4000"
+export LITELLM_API_KEY="sk-1234"
+export LITELLM_MODEL="claude-sonnet-4-20250514"
+```
+
+Or just use the defaults - it'll connect to `http://localhost:4000` by default.
+
+## Example Config File
+
+If you want to use multiple models, create a `config.yaml` (see `config.example.yaml`):
+
+```yaml
+model_list:
+  - model_name: bedrock-claude-sonnet-4
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+      aws_region_name: "us-east-1"
+```
+
+Then start LiteLLM with: `litellm --config config.yaml`
+
+## How It Works
+
+The key is pointing the Agent SDK to LiteLLM instead of directly to Anthropic:
+
+```python
+# Point to LiteLLM gateway (not Anthropic)
+os.environ["ANTHROPIC_BASE_URL"] = "http://localhost:4000"
+os.environ["ANTHROPIC_API_KEY"] = "sk-1234"  # Your LiteLLM key
+
+# Use any model configured in LiteLLM
+options = ClaudeAgentOptions(
+    model="bedrock-claude-sonnet-4",  # or gpt-4, or anything else
+    system_prompt="You are a helpful assistant.",
+    max_turns=50,
+)
+```
+
+Note: Don't add `/anthropic` to the base URL - LiteLLM handles the routing automatically.
+
+## Why Use This?
+
+- **Switch providers easily**: Use the same code with OpenAI, Bedrock, Azure, etc.
+- **Cost tracking**: LiteLLM tracks spending across all your agent conversations
+- **Rate limiting**: Set budgets and limits on your agent usage
+- **Load balancing**: Distribute requests across multiple API keys or regions
+- **Fallbacks**: Automatically retry with a different model if one fails
+
+## Troubleshooting
+
+**Connection errors?**
+- Make sure LiteLLM is running: `litellm --model your-model`
+- Check the URL is correct (default: `http://localhost:4000`)
+
+**Authentication errors?**
+- Verify your LiteLLM API key is correct
+- Make sure the model is configured in your LiteLLM setup
+
+**Model not found?**
+- Check the model name matches what's in your LiteLLM config
+- Run `litellm --model your-model` to test it works
+
+## Learn More
+
+- [LiteLLM Docs](https://docs.litellm.ai/)
+- [Claude Agent SDK](https://github.com/anthropics/anthropic-agent-sdk)
+- [LiteLLM Proxy Guide](https://docs.litellm.ai/docs/proxy/quick_start)
@@ -0,0 +1,25 @@
+model_list:
+  - model_name: bedrock-claude-sonnet-3.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-3-5-sonnet-20240620-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-opus-4.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-opus-4-5-20251101-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-nova-premier
+    litellm_params:
+      model: "bedrock/amazon.nova-premier-v1:0"
+      aws_region_name: "us-east-1"
@@ -0,0 +1,196 @@
+"""
+Simple Interactive Claude Agent SDK CLI using LiteLLM Gateway
+
+This example demonstrates an interactive CLI chat with the Anthropic Agent SDK using LiteLLM as a proxy.
+LiteLLM acts as a unified interface, allowing you to use any LLM provider (OpenAI, Azure, Bedrock, etc.)
+through the Claude Agent SDK by pointing it to the LiteLLM gateway.
+"""
+
+import os
+import asyncio
+import httpx
+from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions
+
+
+class Config:
+    """Configuration for LiteLLM Gateway connection"""
+    
+    # LiteLLM proxy URL (default to local instance)
+    LITELLM_PROXY_URL = os.getenv("LITELLM_PROXY_URL", "http://localhost:4000")
+    
+    # LiteLLM API key (master key or virtual key)
+    LITELLM_API_KEY = os.getenv("LITELLM_API_KEY", "sk-1234")
+    
+    # Model name as configured in LiteLLM (e.g., "bedrock-claude-sonnet-4", "gpt-4", etc.)
+    LITELLM_MODEL = os.getenv("LITELLM_MODEL", "bedrock-claude-sonnet-4.5")
+
+
+async def fetch_available_models(base_url: str, api_key: str) -> list[str]:
+    """
+    Fetch available models from LiteLLM proxy /models endpoint
+    """
+    try:
+        async with httpx.AsyncClient() as client:
+            response = await client.get(
+                f"{base_url}/models",
+                headers={"Authorization": f"Bearer {api_key}"},
+                timeout=10.0
+            )
+            response.raise_for_status()
+            data = response.json()
+            return [model["id"] for model in data.get("data", [])]
+    except Exception as e:
+        print(f"⚠️  Warning: Could not fetch models from proxy: {e}")
+        print("Using default model list...")
+        # Fallback to default models
+        return [
+            "bedrock-claude-sonnet-3.5",
+            "bedrock-claude-sonnet-4",
+            "bedrock-claude-sonnet-4.5",
+            "bedrock-claude-opus-4.5",
+            "bedrock-nova-premier",
+        ]
+
+
+async def interactive_chat():
+    """
+    Interactive CLI chat with the agent
+    """
+    config = Config()
+    
+    # Configure Anthropic SDK to point to LiteLLM gateway
+    # Note: We don't add /anthropic to the base URL - LiteLLM handles routing
+    litellm_base_url = config.LITELLM_PROXY_URL.rstrip('/')
+    os.environ["ANTHROPIC_BASE_URL"] = litellm_base_url
+    os.environ["ANTHROPIC_API_KEY"] = config.LITELLM_API_KEY
+    
+    # Fetch available models from proxy
+    available_models = await fetch_available_models(litellm_base_url, config.LITELLM_API_KEY)
+    
+    current_model = config.LITELLM_MODEL
+    
+    print("=" * 70)
+    print("🤖 Claude Agent SDK with LiteLLM Gateway - Interactive Chat")
+    print("=" * 70)
+    print(f"🚀 Connected to: {litellm_base_url}")
+    print(f"📦 Current model: {current_model}")
+    print("\nType your messages below. Commands:")
+    print("  - 'quit' or 'exit' to end the conversation")
+    print("  - 'clear' to start a new conversation")
+    print("  - 'model' to switch models")
+    print("  - 'models' to list available models")
+    print("=" * 70)
+    print()
+    
+    while True:
+        # Configure agent options for each conversation
+        options = ClaudeAgentOptions(
+            system_prompt="You are a helpful AI assistant. Be concise, accurate, and friendly.",
+            model=current_model,
+            max_turns=50,
+        )
+        
+        # Create agent client
+        async with ClaudeSDKClient(options=options) as client:
+            conversation_active = True
+            
+            while conversation_active:
+                # Get user input
+                try:
+                    user_input = input("\n👤 You: ").strip()
+                except (EOFError, KeyboardInterrupt):
+                    print("\n\n👋 Goodbye!")
+                    return
+                
+                # Handle commands
+                if user_input.lower() in ['quit', 'exit']:
+                    print("\n👋 Goodbye!")
+                    return
+                
+                if user_input.lower() == 'clear':
+                    print("\n🔄 Starting new conversation...\n")
+                    conversation_active = False
+                    continue
+                
+                if user_input.lower() == 'models':
+                    print("\n📋 Available models:")
+                    for i, model in enumerate(available_models, 1):
+                        marker = "✓" if model == current_model else " "
+                        print(f"  {marker} {i}. {model}")
+                    continue
+                
+                if user_input.lower() == 'model':
+                    print("\n📋 Select a model:")
+                    for i, model in enumerate(available_models, 1):
+                        marker = "✓" if model == current_model else " "
+                        print(f"  {marker} {i}. {model}")
+                    
+                    try:
+                        choice = input("\nEnter number (or press Enter to cancel): ").strip()
+                        if choice:
+                            idx = int(choice) - 1
+                            if 0 <= idx < len(available_models):
+                                current_model = available_models[idx]
+                                print(f"\n✅ Switched to: {current_model}")
+                                print("🔄 Starting new conversation with new model...\n")
+                                conversation_active = False
+                            else:
+                                print("❌ Invalid choice")
+                    except (ValueError, IndexError):
+                        print("❌ Invalid input")
+                    continue
+                
+                if not user_input:
+                    continue
+                
+                # Send query to agent with loading indicator
+                print("\n🤖 Assistant: ", end='', flush=True)
+                
+                try:
+                    await client.query(user_input)
+                    
+                    # Show loading indicator
+                    print("⏳ thinking...", end='', flush=True)
+                    
+                    # Stream the response
+                    first_chunk = True
+                    async for msg in client.receive_response():
+                        # Clear loading indicator on first message
+                        if first_chunk:
+                            print("\r🤖 Assistant: ", end='', flush=True)
+                            first_chunk = False
+                        
+                        # Handle different message types
+                        if hasattr(msg, 'type'):
+                            if msg.type == 'content_block_delta':
+                                # Streaming text delta
+                                if hasattr(msg, 'delta') and hasattr(msg.delta, 'text'):
+                                    print(msg.delta.text, end='', flush=True)
+                            elif msg.type == 'content_block_start':
+                                # Start of content block
+                                if hasattr(msg, 'content_block') and hasattr(msg.content_block, 'text'):
+                                    print(msg.content_block.text, end='', flush=True)
+                        
+                        # Fallback to original content handling
+                        if hasattr(msg, 'content'):
+                            for content_block in msg.content:
+                                if hasattr(content_block, 'text'):
+                                    print(content_block.text, end='', flush=True)
+                    
+                    print()  # New line after response
+                    
+                except Exception as e:
+                    print(f"\r\n❌ Error: {e}")
+                    print("Please check your LiteLLM gateway is running and configured correctly.")
+
+
+def main():
+    """Run interactive chat"""
+    try:
+        asyncio.run(interactive_chat())
+    except KeyboardInterrupt:
+        print("\n\n👋 Goodbye!")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,2 @@
+claude-agent-sdk
+httpx>=0.27.0
@@ -76,6 +76,13 @@ BEDROCK_COMPUTER_USE_TOOLS = [
    "text_editor_",
 ]

+# Beta header patterns that are not supported by Bedrock Converse API
+# These will be filtered out to prevent errors
+UNSUPPORTED_BEDROCK_CONVERSE_BETA_PATTERNS = [
+    "advanced-tool-use",  # Bedrock Converse doesn't support advanced-tool-use beta headers
+    "prompt-caching",  # Prompt caching not supported in Converse API
+]
+

 class AmazonConverseConfig(BaseConfig):
    """
@@ -610,6 +617,37 @@ class AmazonConverseConfig(BaseConfig):

        return transformed_tools

+    def _filter_unsupported_beta_headers_for_bedrock(
+        self, model: str, beta_list: list
+    ) -> list:
+        """
+        Remove beta headers that are not supported on Bedrock Converse API for the given model.
+
+        Extended thinking beta headers are only supported on specific Claude 4+ models.
+        Some beta headers are universally unsupported on Bedrock Converse API.
+
+        Args:
+            model: The model name
+            beta_list: The list of beta headers to filter
+
+        Returns:
+            Filtered list of beta headers
+        """
+        filtered_betas = []
+        
+        # 1. Filter out beta headers that are universally unsupported on Bedrock Converse
+        for beta in beta_list:
+            should_keep = True
+            for unsupported_pattern in UNSUPPORTED_BEDROCK_CONVERSE_BETA_PATTERNS:
+                if unsupported_pattern in beta.lower():
+                    should_keep = False
+                    break
+            
+            if should_keep:
+                filtered_betas.append(beta)
+                
+        return filtered_betas
+
    def _separate_computer_use_tools(
        self, tools: List[OpenAIChatCompletionToolParam], model: str
    ) -> Tuple[
@@ -1088,7 +1126,14 @@ class AmazonConverseConfig(BaseConfig):
                if beta not in seen:
                    unique_betas.append(beta)
                    seen.add(beta)
-            additional_request_params["anthropic_beta"] = unique_betas
+            
+            # Filter out unsupported beta headers for Bedrock Converse API
+            filtered_betas = self._filter_unsupported_beta_headers_for_bedrock(
+                model=model,
+                beta_list=unique_betas,
+            )
+            
+            additional_request_params["anthropic_beta"] = filtered_betas

        return bedrock_tools, anthropic_beta_list

@@ -54,6 +54,7 @@ class AmazonAnthropicClaudeMessagesConfig(
    # These will be filtered out to prevent 400 "invalid beta flag" errors
    UNSUPPORTED_BEDROCK_INVOKE_BETA_PATTERNS = [
        "advanced-tool-use",  # Bedrock Invoke doesn't support advanced-tool-use beta headers
+        "prompt-caching-scope"
    ]

    def __init__(self, **kwargs):
@@ -1,101 +1,25 @@
 model_list:
-  - model_name: "*"
+  - model_name: bedrock-claude-sonnet-3.5
    litellm_params:
-      model: "*"
-  - model_name: "gpt-4"
+      model: "bedrock/us.anthropic.claude-3-5-sonnet-20240620-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4
    litellm_params:
-      model: "gpt-4"
-      api_key: os.environ/OPENAI_API_KEY
-  - model_name: "gpt-3.5-turbo"
+      model: "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4.5
    litellm_params:
-      model: "gpt-3.5-turbo"
-      api_key: os.environ/OPENAI_API_KEY
+      model: "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+      aws_region_name: "us-east-1"

-general_settings:
-  master_key: sk-1234
+  - model_name: bedrock-claude-opus-4.5
+    litellm_params:
+      model: "bedrock/converse/us.anthropic.claude-opus-4-5-20251101-v1:0"
+      aws_region_name: "us-east-1"

-# ───────────────────────────────────────────────
-# POLICIES - Define WHAT guardrails to apply
-# ───────────────────────────────────────────────
-#
-# Policies define guardrails with:
-# - inherit: Inherit guardrails from another policy
-# - description: Human-readable description
-# - guardrails.add: Add guardrails (on top of inherited)
-# - guardrails.remove: Remove guardrails (from inherited)
-# - condition.model: Model pattern (exact or regex) for when guardrails apply
-#
-policies:
-  # Global baseline policy
-  global-baseline:
-    description: "Base guardrails for all requests"
-    guardrails:
-      add:
-        - pii_blocker
-
-  # Healthcare policy - inherits from global-baseline
-  healthcare-compliance:
-    inherit: global-baseline
-    description: "HIPAA compliance for healthcare teams"
-    guardrails:
-      add:
-        - hipaa_audit
-
-  # Dev policy - inherits but removes PII blocker for testing
-  internal-dev:
-    inherit: global-baseline
-    description: "Relaxed policy for internal development"
-    guardrails:
-      add:
-        - toxicity_filter
-      remove:
-        - pii_blocker
-
-  # Policy with model condition (regex pattern)
-  gpt4-safety:
-    description: "Extra safety for GPT-4 models"
-    guardrails:
-      add:
-        - toxicity_filter
-    condition:
-      model: "gpt-4.*"  # regex: matches gpt-4, gpt-4-turbo, gpt-4o, etc.
-
-  # Policy with model condition (exact match list)
-  bedrock-compliance:
-    description: "Compliance for Bedrock models"
-    guardrails:
-      add:
-        - strict_pii_blocker
-    condition:
-      model: ["bedrock/claude-3", "bedrock/claude-2"]  # exact matches
-
-# ───────────────────────────────────────────────
-# POLICY ATTACHMENTS - Define WHERE policies apply
-# ───────────────────────────────────────────────
-#
-# Attachments are REQUIRED to make policies active.
-# A policy without an attachment will not be applied.
-#
-policy_attachments:
-  # Global attachment - applies to all requests
-  - policy: global-baseline
-    scope: "*"
-
-  # Team-specific attachment
-  - policy: healthcare-compliance
-    teams:
-      - healthcare-team
-      - medical-research
-
-  # Key pattern attachment
-  - policy: internal-dev
-    keys:
-      - "dev-key-*"
-      - "test-key-*"
-
-  # Model-specific policies (attached globally, condition filters by model)
-  - policy: gpt4-safety
-    scope: "*"
-
-  - policy: bedrock-compliance
-    scope: "*"
+  - model_name: bedrock-nova-premier
+    litellm_params:
+      model: "bedrock/us.amazon.nova-premier-v1:0"
+      aws_region_name: "us-east-1"   
@@ -0,0 +1,120 @@
+"""
+E2E tests for Claude Agent SDK with LiteLLM Proxy using Bedrock models.
+
+Tests streaming messages across different Bedrock models:
+- Regular Bedrock Claude Sonnet 4.5
+- Bedrock Converse Claude Sonnet 4.5
+- AWS Nova Premier
+"""
+
+import os
+import pytest
+import asyncio
+from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions
+
+
+# Test models from proxy_config.yaml
+TEST_MODELS = [
+    ("bedrock-claude-sonnet-4.5", "Bedrock Invoke API"),
+    ("bedrock-converse-claude-sonnet-4.5", "Bedrock Converse API"),
+    ("bedrock-nova-premier", "AWS Nova Premier"),
+]
+
+
+@pytest.fixture(scope="module")
+def litellm_proxy_config():
+    """Configure connection to LiteLLM proxy"""
+    proxy_url = os.getenv("LITELLM_PROXY_URL", "http://localhost:4000")
+    api_key = os.getenv("LITELLM_API_KEY", "sk-1234")
+    
+    # Set environment variables for Claude Agent SDK
+    os.environ["ANTHROPIC_BASE_URL"] = proxy_url.rstrip('/')
+    os.environ["ANTHROPIC_API_KEY"] = api_key
+    
+    return {
+        "proxy_url": proxy_url,
+        "api_key": api_key,
+    }
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name,model_description", TEST_MODELS)
+async def test_claude_agent_sdk_streaming(litellm_proxy_config, model_name, model_description):
+    """
+    Test streaming messages with Claude Agent SDK through LiteLLM proxy.
+    
+    This validates:
+    1. Claude Agent SDK can connect to LiteLLM proxy
+    2. Streaming works correctly
+    3. Different Bedrock models (Invoke, Converse, Nova) work end-to-end
+    """
+    print(f"\n{'='*60}")
+    print(f"Testing: {model_name} ({model_description})")
+    print(f"{'='*60}")
+    
+    # Configure agent options
+    options = ClaudeAgentOptions(
+        system_prompt="You are a helpful AI assistant. Be concise.",
+        model=model_name,
+        max_turns=5,
+    )
+    
+    # Test query
+    test_query = "Say 'Hello from LiteLLM!' and nothing else."
+    
+    # Track streaming
+    received_chunks = []
+    full_response = ""
+    
+    try:
+        async with ClaudeSDKClient(options=options) as client:
+            await client.query(test_query)
+            
+            # Collect streaming response
+            async for msg in client.receive_response():
+                # Handle different message types
+                if hasattr(msg, 'type'):
+                    if msg.type == 'content_block_delta':
+                        # Streaming text delta
+                        if hasattr(msg, 'delta') and hasattr(msg.delta, 'text'):
+                            chunk_text = msg.delta.text
+                            received_chunks.append(chunk_text)
+                            full_response += chunk_text
+                    elif msg.type == 'content_block_start':
+                        # Start of content block
+                        if hasattr(msg, 'content_block') and hasattr(msg.content_block, 'text'):
+                            chunk_text = msg.content_block.text
+                            received_chunks.append(chunk_text)
+                            full_response += chunk_text
+                
+                # Fallback to content handling
+                if hasattr(msg, 'content'):
+                    for content_block in msg.content:
+                        if hasattr(content_block, 'text'):
+                            chunk_text = content_block.text
+                            received_chunks.append(chunk_text)
+                            full_response += chunk_text
+        
+        # Assertions
+        print(f"\n✅ Received {len(received_chunks)} chunks")
+        print(f"📝 Full response: {full_response[:100]}...")
+        
+        # Verify we got a response
+        assert len(full_response) > 0, f"No response received from {model_name}"
+        
+        # Verify streaming (should have multiple chunks for most responses)
+        # Note: Very short responses might come in 1 chunk, so we just verify we got content
+        assert len(received_chunks) > 0, f"No chunks received from {model_name}"
+        
+        # Verify response contains expected content (case insensitive)
+        assert "hello" in full_response.lower(), f"Response doesn't contain expected greeting: {full_response}"
+        
+        print(f"✅ Test passed for {model_name}")
+        
+    except Exception as e:
+        pytest.fail(f"Test failed for {model_name} ({model_description}): {str(e)}")
+
+
+if __name__ == "__main__":
+    # Run tests
+    pytest.main([__file__, "-v", "-s"])
@@ -0,0 +1,31 @@
+model_list:
+  - model_name: bedrock-claude-sonnet-3.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-3-5-sonnet-20240620-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-sonnet-4.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-claude-opus-4.5
+    litellm_params:
+      model: "bedrock/us.anthropic.claude-opus-4-5-20251101-v1:0"
+      aws_region_name: "us-east-1"
+
+  - model_name: bedrock-nova-premier
+    litellm_params:
+      model: "bedrock/amazon.nova-premier-v1:0"
+      aws_region_name: "us-east-1"
+
+  # Converse API models
+  - model_name: bedrock-converse-claude-sonnet-4.5
+    litellm_params:
+      model: "bedrock_converse/us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+      aws_region_name: "us-east-1"