Skip to content

QuickStart: Generate Production-Ready Projects with agentcore create

This guide shows how to use the Amazon Bedrock AgentCore starter toolkit to scaffold complete AgentCore projects—either runtime-only or full monorepos with infrastructure-as-code (IaC).

agentcore create generates a working agent implementation, model client, MCP integration, and optional IaC stacks (CDK or Terraform) that provision AgentCore Runtime, Gateway, Memory, and supporting resources in AWS.

All create projects use the Bedrock AgentCore SDK to define an ASGI app that is deployable to the HTTP protocol on AgentCore runtime.

What You Can Generate

agentcore create supports two high-level project templates.

Runtime-Only Template (--template basic)

Generates:

  • src/ with ready-to-run agent code
  • Model loader wired for your selected provider (Anthropic, Bedrock, OpenAI, Gemini)
  • Built-in function tools
  • Optional MCP client
  • No infrastructure code

Use this template for lightweight deployments, and quick iteration. After creation, agentcore launch will zip your code and deploy an AgentCore runtime using the direct code deployment mode.

Production Template (--template production)

Generates:

  • src/ (agent code)
  • mcp/ (gateway tool Lambda)
  • cdk/ or terraform/ (based on --iac selection)
  • IaC modeling:
  • AgentCore Runtime
  • AgentCore Gateway (MCP mode)
  • Cognito OAuth2 client credentials
  • Memory resource
  • Automatic Dockerfile generation with modeled Docker Container deployment

Use this template for full end-to-end AWS deployments.

Prerequisites

  • Python 3.10+
  • uv installed
  • AWS account with credentials configured
  • For basic template, required permissions defined by, Use the starter toolkit
  • For production template
    • IAM permissions sufficient to deploy generated resources
    • Node.js 18+ for CDK projects
    • Terraform 1.2+ for Terraform projects

Step 1: Create a New Project

Run the CLI in interactive mode:

agentcore create

You will be prompted for:

  • Project name
  • Agent SDK (AutoGen, CrewAI, LangGraph, Strands, etc.)
  • Template (basic or production)
  • IaC provider (CDK or Terraform, if applicable)
  • Model provider
  • Whether to include MCP integration
  • Whether to include memory
  • Whether to load defaults from .bedrock_agentcore.yaml

Optional: Run agentcore create

For production templates you will be prompted whether to run agentcore create first. This lets you predefine authorization, headers, memory configuration, and agent details.

Step 2: Inspect the Generated Project

Your output layout depends on the selected template.

Basic Template

my_project/
  src/
    main.py
    model/
    mcp_client/
  .bedrock_agentcore.yaml
  README.md

Includes:

  • Entrypoint (main.py) for local or direct runtime hosting
  • Model loader
  • Optional MCP tools
  • Function tools depending on selected SDK

Production Template (IaC + Runtime)

my_project/
  src/
  mcp/
    lambda/handler.py
  cdk/      OR     terraform/
  .bedrock_agentcore.yaml
  README.md

Includes:

  • Agent runtime code
  • Gateway Lambda used as an MCP target

  • Full IaC modeling:

  • Runtime + endpoints

  • Gateway (MCP)
  • Cognito OAuth2
  • Memory
  • Network + environment variables
  • Container packaging config

Step 3: Local Development

Create and activate a virtual environment:

cd src
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
cd ..

Start the local dev server:

agentcore dev

Invoke it from another terminal:

agentcore invoke --dev '{"prompt": "What can you do?"}'

Hot reload is enabled automatically.

Step 4: Deploy

Basic template

Deploy the basic template with agentcore launch.

If you wish to further configure your project, first run agentcore configure

Production Template

Production Ready Checklist

Before using your generated project in a production environment, consult the following checklist:

  • Security: Ensure secrets and API keys are properly handled. AgentCore Identity or AWS Secrets Manager are secure managed solutions.
  • Build Environment: Confirm Docker builds are being executed in the desired environment. This template uses local Docker builds by default. Consider AWS CodeBuild.
  • Observability: After deploying, enable AgentCore observability to allow OpenTelemetry span data to be published to AWS CloudWatch.
  • CI/CD: Build your new project into a CI/CD pipeline to achieve automated builds, rollbacks, and multiple deployment environments. Consider AWS CodePipeline.
  • Access Control: Configure access for clients to call into your AgentCore Runtime. Take advantage of the multiple endpoints (DEFAULT, PROD, DEV) created by this template.
  • Testing Write unit tests in the generated test/ directory. Implement E2E tests for further coverage.
  • Error Handling Implement graceful and consistent error handling logic throughout your code.

CDK

cd cdk
npm install
npm run cdk synth
npm run cdk:deploy

Make sure Node 18+ is installed.

Terraform

cd terraform
terraform init
terraform plan   # optional
terraform apply

Make sure Terraform 1.2+ is installed.

Step 5: Test Your Deployed Agent

After deploy completes:

agentcore status

When all resources show active, invoke the deployed agent:

agentcore invoke '{"prompt": "Tell me a joke"}'

Step 6: Clean Up

agentcore destroy

Or, for production template, delete stacks and resources using CDK/Terraform directly.

Additional Notes

Model Provider Authentication

  • Bedrock clients use IAM automatically

  • Third-party providers use:

  • AgentCore Identity in deployed environments

  • .env.local fallback in local dev (LOCAL_DEV=1)

For the production template, it is your responsibility to implement API key handling. Using Bedrock AgentCore Identity or AWS Secrets Manager is recommended.

MCP Tools

Generator output will provide the correct MCP adapter for your selected SDK, such as:

  • AutoGen Streamable HTTP MCP adapter
  • CrewAI MCP adapter
  • Gateway-integrated MCP Lambda target

These are included automatically based on your selections.

For the production template, a custom MCP tool is defined in mcp/lambda/handler.py.

A2A and MCP Protocols

MCP, and A2A the other two protocols supported by the AgentCore Runtime Service Contract, are not currently supported by the create tool out of the box. Adapting a create output for another protocol can also be considered.

Next Steps

  • Customize agent logic in src/main.py
  • Add additional MCP integrations in src/mcp_client/
  • For production template, modify your project to adhere to the production ready checklist.
  • Ensure that src/model/load.py has your desired LLM provider configuration.