Skip to content
CrewAI Agent Starter AI Agents · CrewAI 1.15.28 Launch on AWS

Getting started

Launch the CrewAI Agent Starter, run the two-agent sales task, and check the same result after a reboot.

You need: an AWS account that can subscribe on AWS Marketplace, launch EC2, and call Amazon Bedrock in us-east-1 if you want the real-model demo. Session Manager needs an instance profile and the Session Manager plugin on your workstation. SSH is optional.

Watch on YouTube.

1. Subscribe

Open the CrewAI Agent Starter listing, choose View purchase options, review the terms, and subscribe. When the subscription is active, choose Launch your software.

2. Launch the instance

On the EC2 launch page:

  1. Choose t3.medium (4 GiB of memory) or larger.
  2. Under Configure storage, use an encrypted gp3 root volume of at least 24 GiB.
  3. Require IMDSv2, and leave IMDSv2 reachable by the ubuntu user. The starter reads temporary role credentials from it and does not store AWS keys.
  4. For Session Manager, attach an instance profile with AmazonSSMManagedInstanceCore. Add the Bedrock permission in step 4 before the real-model demo.
  5. Leave application ports closed. If you also want SSH, allow TCP 22 only from your source address.
  6. Choose Launch instance.

3. Connect

Session Manager starts as ssm-user. From the EC2 console, choose Connect, then Session Manager, then Connect. Or:

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

Wait for first boot. The initialization unit is a one-shot service. active (exited) is the successful state.

sudo cloud-init status --wait
sudo systemctl status sushi-crewai-firstboot --no-pager

That unit creates /home/ubuntu/crewai-workspace/ and /home/ubuntu/crewai-results/ with mode 0700, plus /etc/crewai/. A reboot keeps them.

4. Allow Amazon Bedrock

The two agents call Amazon Nova Lite in us-east-1 when you run the default demo. Each run makes 2 to 8 model calls, and model charges apply.

  • Allow bedrock:InvokeModel on arn:aws:bedrock:us-east-1::foundation-model/amazon.nova-lite-v1:0.
  • Allow HTTPS egress to Amazon Bedrock in us-east-1, and to the Session Manager endpoints if you use Session Manager.
  • Enable Amazon Nova Lite in the account if it is not already enabled.

The starter does not read AWS keys from disk. The offline fixture in Configuration skips the provider.

5. Run the two-agent task

sudo -iu ubuntu bash /opt/sushi/docs/demo.sh

The analyst calls sum_sales_csv once. The tool reads this run's six-row synthetic CSV and totals integer cents by region. The reporter copies those findings into JSON. Separate code adds the grand total from the saved CSV. A different process recalculates every field without importing CrewAI or calling a model.

Each run writes a new directory under /home/ubuntu/crewai-results/ and prints its manifest path. The directory holds seven files: sales.csv, result.json, provider-trace.json, tool-result.json, tool-step.json, task-outputs.json, and manifest.json.

A real model run uses Bedrock native tool use and a version 2 manifest. That manifest binds the original provider responses, the calculator result, the CrewAI after-tool-call receipt, and both task outputs. Readback rejects a skipped tool or a changed finding. The dataset identity and the amounts change from run to run. A failed run keeps its directory. Fix the reported cause before you start another one.

6. Read the result again

Save the directory the demo printed. After a reboot, reconnect and run this as ubuntu. It checks that directory again and does not call the model.

sudo -iu ubuntu
read -r -p 'Saved demo directory: ' result_dir
sushi-crewai read "$result_dir" --require-provider

--require-provider accepts the Bedrock manifest and rejects an offline fixture. No web service has to be restarted.

Image notes are /opt/sushi/docs/README.md and the published getting-started page linked from the listing. Next, see Configuration for the offline fixture.

Last updated October 27, 2026 · Suggest an improvement