Forem: Haowen Huang

Running Claude Code and Claude Desktop on Amazon Bedrock

Haowen Huang — Mon, 04 May 2026 06:35:23 +0000

Abstract: This article walks through configuring both Claude Code (terminal CLI) and Claude Desktop (Cowork) to use Amazon Bedrock as the inference backend — no Anthropic API key or subscription required. Claude Code needs two environment variables in ~/.claude/settings.json. Claude Desktop needs a few fields in the built-in Setup UI. Both share the same AWS credentials and Bedrock model access. The entire setup, including troubleshooting, was done through AI-assisted development with Kiro.

In a previous post, we covered how to run the Claude Agent SDK (Python) on Amazon Bedrock with just two environment variables. This post continues the series, covering the other two main ways developers interact with Claude daily: the terminal CLI and the desktop app.

If you're already in the AWS ecosystem, routing Claude through Amazon Bedrock keeps your data in your own account, bills through your existing AWS agreement, and eliminates the need for a separate Anthropic subscription. This post covers two scenarios:

Scenario A: Claude Code CLI — the terminal-based coding tool
Scenario B: Claude Desktop (Cowork 3P) — the GUI desktop app

Both use the same AWS IAM user and Bedrock model access.

Prerequisites (Shared)

Before either scenario, make sure you have:

AWS CLI installed and credentials configured

aws --version
aws sts get-caller-identity

Bedrock model access enabled — Serverless foundation models on Amazon Bedrock are now automatically enabled when first invoked. However, for Anthropic models, first-time users may need to submit use case details before access is granted. Simply select a model from the Model catalog and try invoking it — you'll be prompted if any additional steps are needed.
IAM permissions — Your IAM user or role needs:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream",
        "bedrock:ListInferenceProfiles",
        "bedrock:GetInferenceProfile"
      ],
      "Resource": [
        "arn:aws:bedrock:*:*:inference-profile/*",
        "arn:aws:bedrock:*:*:application-inference-profile/*",
        "arn:aws:bedrock:*:*:foundation-model/*"
      ]
    }
  ]
}

Know your inference profile IDs — Bedrock requires cross-region inference profile IDs (with a region prefix like us.), not bare model IDs. The ones used in this post:

Model	Inference Profile ID
Claude Opus 4.7	`us.anthropic.claude-opus-4-7`
Claude Opus 4.6	`us.anthropic.claude-opus-4-6-v1`

Scenario A: Claude Code CLI

Claude Code is Anthropic's terminal-based AI coding tool. Install it via npm:

npm install -g @anthropic-ai/claude-code

Step 1: Edit settings.json

The cleanest approach is writing the Bedrock configuration into Claude Code's settings file, so you don't need to export environment variables every time.

Edit ~/.claude/settings.json:

{
  "model": "us.anthropic.claude-opus-4-6-v1",
  "env": {
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "AWS_REGION": "us-east-1"
  }
}

Two key fields:

CLAUDE_CODE_USE_BEDROCK: Tells Claude Code to use Bedrock instead of the Anthropic API.
AWS_REGION: Required. Claude Code does not read this from ~/.aws/config.

The model field uses the Bedrock inference profile ID. Without it, Claude Code picks its own default.

Step 2: Run and Verify

claude

Inside Claude Code, run /status. You should see:

API provider: Amazon Bedrock
AWS region: us-east-1
Model: us.anthropic.claude-opus-4-6-v1

That's it. Claude Code picks up your AWS credentials through the standard SDK credential chain (~/.aws/credentials, environment variables, IAM role, SSO profile).

Switching Models

There are two ways to switch models:

At runtime — Inside Claude Code, use the /model command:

/model us.anthropic.claude-opus-4-7

This switches immediately without restarting. Use /model us.anthropic.claude-opus-4-6-v1 to switch back.

Permanently — Change the model field in ~/.claude/settings.json:

{
  "model": "us.anthropic.claude-opus-4-7",
  "env": {
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "AWS_REGION": "us-east-1"
  }
}

Restart Claude Code (/exit, then claude) for the change to take effect.

Optional: Install Skills

Claude Code supports Skills — packaged instructions that extend Claude's capabilities. Anthropic publishes an official set at github.com/anthropics/skills.

To install them, run inside Claude Code:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills
/reload-plugins

This gives you skills like mcp-builder, webapp-testing, claude-api, document generation (docx/pdf/pptx/xlsx), and more.

After installation, your ~/.claude/settings.json will be updated automatically with the new plugin entries:

{
  "model": "us.anthropic.claude-opus-4-6-v1",
  "env": {
    "CLAUDE_CODE_USE_BEDROCK": "1",
    "AWS_REGION": "us-east-1"
  },
  "enabledPlugins": {
    "document-skills@anthropic-agent-skills": true,
    "example-skills@anthropic-agent-skills": true
  },
  "extraKnownMarketplaces": {
    "anthropic-agent-skills": {
      "source": {
        "source": "github",
        "repo": "anthropics/skills"
      }
    }
  }
}

Compared to the base configuration in Step 1, two new sections are added:

enabledPlugins: Tracks which Skills are active.
extraKnownMarketplaces: Records the GitHub source for the Skills marketplace.

Example: Turning a Blog Post into a Slide Deck

Once Skills are installed, you can use them directly in Claude Code. For example, I previously wrote a blog post "Deploying OpenClaw on Amazon EC2 - A Developer's Perspective", covering one-click CloudFormation deployment, VPC Endpoints security architecture, and multi-region inference profile configuration. It's a detailed post. If I needed to present this topic at a developer conference the next day, building slides manually would take too long. With the pptx skill, it takes one command:

❯ /example-skills:pptx
⏺ PPTX skill loaded. What presentation would you like me to create?

❯ Turn the following blog post into a slide deck:
  https://builder.aws.com/content/3CQFzc91nM9sGkfHRbpjo4eXojf/deploying-openclaw-on-amazon-ec2-a-developers-perspective

Claude Code reads the blog, extracts the architecture descriptions, deployment steps, and key configurations, organizes them into well-structured slides, and generates a .pptx file — all within the terminal, powered by Amazon Bedrock. From blog post to presentation-ready slides in a single command.

Scenario B: Claude Desktop (Cowork 3P)

Claude Desktop can run in "third-party platform" mode, using Bedrock for inference instead of Anthropic's own servers. This is called Cowork 3P.

Step 1: Open the Setup UI

Open Claude Desktop (you don't need to log in)
Go to Help → Troubleshooting → Enable Developer mode
Go to Developer → Configure third-party inference

This opens the built-in configuration UI.

Step 2: Configure Bedrock

Fill in the following fields:

Field	Value
Inference provider	Bedrock (select the radio button)
AWS region	`us-east-1`
AWS bearer token	Leave empty
Bedrock base URL	Leave empty
AWS profile name	`default`
AWS config directory	Leave empty (uses `~/.aws` by default)

Under IDENTITY & MODELS:

Field	Value
Model list	`us.anthropic.claude-opus-4-6-v1`

The model list uses Bedrock inference profile IDs. You can add multiple models — click + Add for each one. The first entry becomes the default.

Step 3: Apply and Relaunch

Click Apply locally at the bottom of the Setup UI
Fully quit Claude Desktop (Cmd+Q on macOS)
Reopen Claude Desktop

After relaunch, Claude Desktop shows the Bedrock login screen — no Anthropic account needed, just your AWS credentials:

Step 4: Verify

After relaunch, you should see:

Bottom-left corner shows Cowork 3P | Bedrock
The model picker shows us.anthropic.claude-opus-4-6-v1
Claude responds and identifies itself as running on the Bedrock model

Send a test message like "What version of the model are you currently using?" — Claude should confirm it's powered by claude-opus-4-6-v1 in Cowork mode.

Adding More Models Later

To add models (e.g., Opus 4.7) after initial setup:

In the macOS menu bar, go to Help → Troubleshooting → Enable Developer mode
A Developer menu appears in the menu bar — click Developer → Configure third-party inference
Under Model list, click + Add and enter us.anthropic.claude-opus-4-7
Click Apply locally, then Cmd+Q and reopen Claude Desktop

The model picker at the bottom of the chat window will now show all configured models.

How It Works Under the Hood

When you set AWS profile name to default, Claude Desktop reads your ~/.aws/credentials and ~/.aws/config files to authenticate with Bedrock. The three authentication methods are:

Method	When to use
AWS bearer token	If you have a Bedrock API key
AWS profile name	If you have standard AWS credentials (this is what we used)
Credential helper	For enterprise SSO setups with rotating credentials

You only need one of the three. The Setup UI validates the configuration and writes it to ~/Library/Application Support/Claude-3p/claude_desktop_config.json on macOS.

Important: Use the Setup UI

During this setup, I first tried manually editing the claude_desktop_config.json file. This resulted in a persistent "Your provider setup needs a fix" error. The Setup UI handles the correct format and field validation automatically — use it instead of hand-editing the JSON.

Troubleshooting

"on-demand throughput isn't supported" — You're using a bare model ID instead of an inference profile ID. Use the us. prefixed IDs (e.g., us.anthropic.claude-opus-4-6-v1).

"Your provider setup needs a fix" in Claude Desktop — Don't hand-edit the config JSON. Use the Setup UI (Developer → Configure third-party inference) and click Apply locally.

Claude Code /status doesn't show Bedrock — Make sure CLAUDE_CODE_USE_BEDROCK is set to "1" in ~/.claude/settings.json under the env key. Restart Claude Code after changing settings.

AWS credentials not found — Claude Code uses the standard AWS SDK credential chain. Verify with aws sts get-caller-identity. For Claude Desktop, make sure the AWS profile name matches a profile in ~/.aws/credentials.

Summary

	Claude Code CLI	Claude Desktop (Cowork 3P)
Config location	`~/.claude/settings.json`	Setup UI → Apply locally
Key setting	`CLAUDE_CODE_USE_BEDROCK=1`	`inferenceProvider: bedrock`
Auth method	AWS SDK credential chain	AWS profile name
Model format	Inference profile ID	Inference profile ID
Verify	`/status` in Claude Code	Bottom-left shows "Cowork 3P \| Bedrock"

Both scenarios use the same AWS credentials and Bedrock model access. Once you have one working, the other is straightforward.

A Note on the Setup Process

This entire configuration — both scenarios, including troubleshooting the Claude Desktop JSON format issue — was completed through conversation with Kiro, an AI-powered IDE. Kiro checked the AWS environment, verified IAM permissions, wrote the settings files, researched the official documentation, and diagnosed the Setup UI requirement for Claude Desktop — all without leaving the IDE.

Tested with Claude Code 2.1.123, Claude Desktop (Cowork 3P), Claude Opus 4.6 & 4.7 on Amazon Bedrock, us-east-1

Deploying OpenClaw on Amazon EC2 - A Developer's Perspective

Haowen Huang — Thu, 16 Apr 2026 04:10:44 +0000

What is OpenClaw?

OpenClaw is a personal AI assistant that you can deploy on your own infrastructure. It can respond to you through channels you already use — WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, WeChat, Lark, and 20+ other platforms. It can have voice conversations on macOS/iOS/Android, and render a real-time Canvas you can control. The Gateway is just the control plane — the product itself is the assistant.

If you want a self-hosted, fast, always-on single-user personal assistant, OpenClaw is it.

Why Run Your Own AI Assistant on AWS?

Most people interact with AI through hosted services. That works — until you care about:

Privacy: Your conversations, code, and data stay in your AWS account
Cost control: Pay only for Bedrock API usage, no per-seat SaaS fees
Customization: Full control over models, plugins, and integrations
Security: No open ports, no public endpoints, SSM-only access

By deploying OpenClaw on AWS with Amazon Bedrock, you can use the latest Claude models (Opus 4.6, Sonnet 4.6) as your personal AI assistant — not just for coding, but for automation, messaging, device control, and more.

I decided to try it out myself. Here's how it went.

Architecture: What You're Deploying

🚀 Quick Start: If you're an experienced developer, you can use the deployment template directly:
curl -O https://raw.githubusercontent.com/hanyun2019/openclaw-on-aws/main/openclaw-deployment.yaml

aws cloudformation create-stack \
  --stack-name openclaw \
  --template-body file://openclaw-deployment.yaml \
  --capabilities CAPABILITY_IAM \
  --region us-east-1
📦 Template URL: github.com/hanyun2019/openclaw-on-aws

⚠️ Non-US Region Users: Additional configuration is required after deployment. See the Non-US Region Deployment Guide below.

Before we dive in, here's what the CloudFormation template sets up:

User Access: Connect via SSM Session Manager to establish a secure tunnel, forwarding local localhost:18789 to the EC2 instance — no inbound ports required.

Key design decisions:

EC2 in the public subnet — for initial setup (package downloads via Internet Gateway)
VPC Endpoints in the private subnet — ensures Bedrock API calls never leave the AWS network
No inbound ports open — all access through SSM port forwarding
Graviton (ARM) instances — better price-performance ratio

Prerequisites

Before you start, make sure you have:

An AWS account with billing enabled
IAM permissions: Ability to create CloudFormation, EC2, IAM, and VPC resources
Local tools installed:
- AWS CLI v2
- SSM Session Manager Plugin

US Region Deployment Guide

If you're deploying in us-east-1 or us-west-2, it's genuinely a one-click experience.

Step 1: Download the Template

curl -O https://raw.githubusercontent.com/hanyun2019/openclaw-on-aws/main/openclaw-deployment.yaml

Step 2: Deploy the Stack

aws cloudformation create-stack \
  --stack-name openclaw \
  --template-body file://openclaw-deployment.yaml \
  --capabilities CAPABILITY_IAM \
  --region us-east-1

Optional parameters:

Parameter	Default	Description
`InstanceType`	`t4g.large`	Graviton instance type (2 vCPU, 8GB RAM)
`KeyPairName`	(none)	EC2 key pair for SSH access (optional)
`CreateVPCEndpoints`	`true`	Create VPC endpoints for private Bedrock access
`AllowedSSHCIDR`	(empty)	CIDR block for SSH access (optional)

For example, to use a larger instance:

aws cloudformation create-stack \
  --stack-name openclaw \
  --template-body file://openclaw-deployment.yaml \
  --capabilities CAPABILITY_IAM \
  --region us-east-1 \
  --parameters ParameterKey=InstanceType,ParameterValue=t4g.xlarge

Step 3: Wait (~10-15 minutes)

aws cloudformation wait stack-create-complete \
  --stack-name openclaw --region us-east-1

Go grab a coffee. ☕

Step 4: Get Your Access Info

aws cloudformation describe-stacks \
  --stack-name openclaw \
  --query 'Stacks[0].Outputs' \
  --region us-east-1

This gives you the instance ID and access URL.

Step 5: Connect via SSM Port Forwarding

# Get Instance ID
INSTANCE_ID=$(aws cloudformation describe-stacks \
  --stack-name openclaw \
  --query 'Stacks[0].Outputs[?OutputKey==`InstanceId`].OutputValue' \
  --output text --region us-east-1)

# Start the tunnel
aws ssm start-session \
  --target $INSTANCE_ID \
  --region us-east-1 \
  --document-name AWS-StartPortForwardingSession \
  --parameters '{"portNumber":["18789"],"localPortNumber":["18789"]}'

Step 6: Open OpenClaw

Navigate to the Step3AccessURL from the stack outputs:

http://localhost:18789/?token=<your-token>

That's it. You're in. 🎉

Non-US Region Deployment Guide

If you're deploying outside the US (e.g., Sydney ap-southeast-2), you need to modify the model IDs.

Note: If you're deploying in us-east-1 or us-west-2, you can skip this section. The template's default us. prefixed model IDs work out of the box in US regions.

The Issue

The template defaults to us.anthropic.claude-opus-4-6-v1 — this is the US region Inference Profile. In other regions, you need to use the corresponding regional prefix.

Solution Steps

1. Connect to the Instance via SSM

aws ssm start-session --target <instance-id> --region ap-southeast-2

2. Query Available Inference Profiles (Optional)

aws bedrock list-inference-profiles --region ap-southeast-2 \
  --query 'inferenceProfileSummaries[*].inferenceProfileId' \
  --output table

3. Update OpenClaw Configuration

nano /home/ubuntu/.openclaw/openclaw.json

Replace the us. prefix in model IDs with your region's prefix:

{
  "models": {
    "providers": {
      "amazon-bedrock": {
        "models": [
          {
            "id": "au.anthropic.claude-opus-4-6-v1",
            "name": "Claude Opus 4.6"
          },
          {
            "id": "au.anthropic.claude-sonnet-4-6",
            "name": "Claude Sonnet 4.6"
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "amazon-bedrock/au.anthropic.claude-opus-4-6-v1"
      }
    }
  }
}

nano save: Ctrl+O → Enter to confirm → Ctrl+X to exit

4. Restart the Gateway

systemctl --user restart openclaw-gateway.service

Done! 🎉

Understanding Inference Profile Prefixes

Here's a quick reference:

Prefix	Scope	Example Regions
`us.`	US cross-region	us-east-1, us-east-2, us-west-2, ca-central-1
`eu.`	Europe cross-region	eu-central-1, eu-west-1, eu-west-3, eu-north-1
`apac.`	Asia-Pacific cross-region	ap-northeast-1, ap-southeast-1, ap-southeast-2, ap-south-1
`au.`	Australia	ap-southeast-2, ap-southeast-4
`jp.`	Japan	ap-northeast-1, ap-northeast-3
`global.`	Global cross-region (routes to optimal commercial region)	Can be called from all commercial regions

Rule of thumb: Use the most specific prefix for your deployment region.

Deploying in us-east-1 or us-west-2? The default us.anthropic.claude-opus-4-6-v1 just works. No changes needed.
Deploying in ap-southeast-2 (Sydney)? Use au.anthropic.claude-opus-4-6-v1.
Deploying in ap-northeast-1 (Tokyo)? Use jp.anthropic.claude-sonnet-4-6 or apac. prefix.
Deploying in eu-west-1 (Ireland)? Use eu.anthropic.claude-opus-4-6-v1.
Not sure? Query aws bedrock list-inference-profiles in your region to see what's available.

Note: Model availability varies by region. For example, Claude Opus 4.6 in Japan may only be available with the global. prefix, while Sonnet 4.6 has a jp. prefix. Use aws bedrock list-inference-profiles to check actual available profiles in your region.

Cost Estimate

Resource	Spec	Est. Monthly Cost	Calculation
EC2 (`t4g.large`)	2 vCPU, 8GB RAM	~$49	$0.0672/hr × 730 hrs
EBS (`gp3`)	50GB	~$4	$0.08/GB-mo × 50GB
VPC Endpoints (×4)	Interface type	~$29	$0.01/hr × 730 hrs × 4 endpoints
Bedrock API	Usage-based	Variable	—
Total (excl. API)	—	~$82	—

Pricing References (us-east-1 region, 2024):

EC2 pricing: AWS EC2 T4g Instances

EBS pricing: AWS EBS Pricing

VPC Endpoint pricing: AWS PrivateLink Pricing

Note: Actual prices may vary by region and time. Please refer to official AWS pricing pages for current rates.

Cost-saving tip: Set CreateVPCEndpoints=false to save ~$29/month. In this case, Bedrock API calls will route through the Internet Gateway to Bedrock's public endpoints. Traffic is still encrypted via TLS, but traverses the public internet rather than staying within the AWS backbone. For scenarios with strict security or compliance requirements, keep VPC endpoints enabled to ensure traffic never leaves the AWS network.

FAQ

How do I access EC2 via SSM and use OpenClaw TUI?

Option 1: Via AWS Console

Open the EC2 Console
Select your OpenClaw instance
Click Connect → Select Session Manager → Click Connect

Option 2: Via AWS CLI

aws ssm start-session --target <instance-id> --region <your-region>

After connecting, switch to the ubuntu user and launch TUI:

sudo su - ubuntu
openclaw tui

How do I configure OpenClaw and switch Bedrock models?

Use the openclaw config command for configuration management, including switching models on Amazon Bedrock:

openclaw config

Stack deployment timed out (WaitCondition failure)?

Connect to the instance via SSM and check the setup log:

cat /var/log/openclaw-setup.log

Common causes:

Network issues during package downloads
Bedrock model access not yet approved (it can take a few minutes after requesting)

How do I retrieve my Gateway token?

# Replace <stack-name> with your stack name
aws ssm get-parameter \
  --name "/openclaw/<stack-name>/gateway-token" \
  --with-decryption \
  --query 'Parameter.Value' \
  --output text

How do I switch models?

Edit agents.defaults.model.primary in /home/ubuntu/.openclaw/openclaw.json, then restart:

systemctl --user restart openclaw-gateway.service

How do I update OpenClaw?

SSH or SSM into the instance and run:

npm update -g openclaw
systemctl --user restart openclaw-gateway.service

How do I tear it all down?

aws cloudformation delete-stack --stack-name openclaw --region us-east-1

This removes everything — EC2 instance, VPC, IAM roles, all of it. Clean.

Wrapping Up

Deploying OpenClaw on AWS is straightforward — with one note about region selection:

US regions (us-east-1, us-west-2): Truly one-click. The template works out of the box.
Other regions: You need to replace the model IDs with the corresponding regional inference profiles.

Either way, the CloudFormation template handles 90% of the work. What you get at the end is a private, secure AI assistant running on your own infrastructure — no data leaving your AWS account, no third-party services in the loop, and the full power of Claude at your fingertips.

Resources:

📦 Deployment Template (GitHub)

Series Preview: Four Ways to Deploy OpenClaw on AWS

This post is the first in the OpenClaw on AWS series. There are four ways to deploy OpenClaw on AWS, each suited for different scenarios:

Deployment Method	Use Case	Status
EC2 Deployment	Full control, custom configuration, persistent runtime	✅ This post
Lightsail Deployment	Simple and fast, fixed monthly cost, beginner-friendly	📝 Coming soon
AgentCore Deployment	Managed service, no infrastructure management	📝 Coming soon
EKS Deployment	Containerized, scalable, enterprise-grade	📝 Coming soon

Stay tuned for upcoming posts!

Industry Trends: The Rise of AI Agents

Beyond OpenClaw, the AI Agent space is evolving rapidly. More developers are exploring how to build autonomous, controllable AI agent systems.

One project worth watching is Hermes Agent — an open-source AI Agent framework developed by Nous Research that has been gaining increasing attention in the developer community. It offers a different approach to building and deploying AI agents.

I'll share hands-on experience with Hermes Agent in upcoming blog posts — stay tuned.

Happy Deploying! 🦞

References

Running Claude Agent SDK with Skills on Amazon Bedrock

Haowen Huang — Wed, 11 Mar 2026 08:56:47 +0000

by Haowen Huang (LinkedIn)

The Claude Agent SDK enables building multi-agent systems with Skills, MCP servers, and subagents. By default, it requires an Anthropic API key. This post shows how to run it entirely on Amazon Bedrock instead—no Anthropic API key needed.

The Problem

The DeepLearning.AI course "Agent Skills with Anthropic" assumes you have an ANTHROPIC_API_KEY. But what if you're already using AWS and want to leverage your existing Amazon Bedrock access?

The Solution: Two Lines of Code

The Claude Agent SDK uses Claude Code under the hood, which supports Amazon Bedrock via environment variables. The key configuration is surprisingly simple:

import os

# Configure Claude Code to use Amazon Bedrock
os.environ["CLAUDE_CODE_USE_BEDROCK"] = "1"
os.environ["AWS_REGION"] = "us-west-2"  # Your preferred region

That's it. Set these before creating ClaudeSDKClient, and the SDK will use your AWS credentials instead of an Anthropic API key.

Prerequisites

Before running, verify your AWS setup:

# Check AWS CLI
aws --version

# Verify credentials
aws configure get aws_access_key_id

# Test Bedrock access
aws bedrock list-foundation-models --query "modelSummaries[?contains(modelId, 'claude')]" --output table

You need:

AWS credentials configured (~/.aws/credentials or environment variables)
Amazon Bedrock model access enabled in your AWS console
IAM permissions for bedrock:InvokeModel

Complete Example

Here's the modified agent.py for Amazon Bedrock:

import asyncio
import os
from dotenv import load_dotenv
from claude_agent_sdk import (
    AgentDefinition, ClaudeSDKClient, ClaudeAgentOptions, AssistantMessage,
)

load_dotenv()
AWS_REGION = os.environ.get("AWS_REGION", "us-west-2")

# Key configuration for Bedrock
os.environ["CLAUDE_CODE_USE_BEDROCK"] = "1"
os.environ["AWS_REGION"] = AWS_REGION

async def main():
    agents = {
        "docs_researcher": AgentDefinition(
            description="Finds information from official documentation.",
            prompt="You research official docs.",
            tools=["WebSearch", "WebFetch"],
            model="haiku"
        ),
    }

    options = ClaudeAgentOptions(
        system_prompt="You are a helpful assistant.",
        allowed_tools=["Skill", "Task", "Write", "Bash", "WebSearch", "WebFetch"],
        model="sonnet",
        agents=agents
    )

    async with ClaudeSDKClient(options=options) as client:
        await client.query("Hello!")
        async for message in client.receive_response():
            if isinstance(message, AssistantMessage):
                print(message.content)

asyncio.run(main())

Adding MCP Servers (Optional)

MCP servers like Notion work the same way. Just ensure the token exists before adding the config:

NOTION_TOKEN = os.environ.get("NOTION_TOKEN")

mcp_servers = {}
if NOTION_TOKEN:
    mcp_servers["notion"] = {
        "command": "npx",
        "args": ["-y", "@notionhq/notion-mcp-server"],
        "env": {"NOTION_TOKEN": NOTION_TOKEN},
    }

options = ClaudeAgentOptions(
    mcp_servers=mcp_servers,
    # ... other options
)

What Works on Amazon Bedrock

Everything from the original SDK works:

Skills (e.g., learning-a-tool)
Subagents with parallel execution
MCP servers (Notion, etc.)
All built-in tools (WebSearch, Bash, Write, etc.)

A Note on the Debugging Process

This entire migration—from identifying the Bedrock configuration, fixing MCP server issues, to successfully running Skills with subagents—was completed through conversation with Kiro, an AI-powered IDE. No external documentation was consulted. Kiro diagnosed issues in real-time, suggested fixes, and validated the setup by running test scripts directly.

The workflow:

Asked Kiro how to use Bedrock instead of Anthropic API
Kiro checked the environment, identified missing dependencies
Kiro found the correct configuration (environment variables, not client parameters)
Fixed runtime errors (MCP config validation when token was missing)
Verified everything worked end-to-end

This demonstrates how AI-assisted development can accelerate debugging unfamiliar SDKs without documentation diving.

Conclusion

Running Claude Agent SDK on Amazon Bedrock requires just two environment variables. Your existing AWS credentials handle authentication, and all SDK features—Skills, subagents, MCP servers—work unchanged.

If you're already in the AWS ecosystem, this is the simplest path to using Claude Agent SDK without managing a separate Anthropic API key.