Claude Managed Agents on AWS Lambda MicroVMs
This reference solution demonstrates how to run Claude's self-hosted sandbox tool execution on AWS Lambda MicroVMs. It implements an orchestrator plus ephemeral-MicroVM-per-session pattern, where each Claude session gets its own Firecracker-isolated compute environment with VM-level isolation and no state shared between sessions.
The control plane is webhook-driven. When a Claude session starts, Anthropic sends a webhook to an Amazon API Gateway endpoint (protected by AWS WAF) that triggers a Lambda launcher. The launcher verifies the signature using a secret from Systems Manager Parameter Store and launches a single MicroVM. The MicroVM fetches per-session credentials, claims the session, executes tools, and posts results back.
The security model keeps organization API keys away from AWS compute. Only per-session references and environment-key parameter names are forwarded, secrets live in Parameter Store, and IAM execution roles enforce least-privilege access. The event-driven design launches fresh MicroVMs on demand, so agents can run untrusted code in strong isolation while preserving full organizational governance and CloudWatch monitoring.
< Back to all repos
Clone this repo
git clone https://github.com/aws-samples/sample-lambda-microvm-claude-managed-agentsAdditional resources
Created by:
