The cloud plumbing that lets AI agents run code without trashing your laptop

What it does

This is the production infrastructure that powers E2B, a service that spins up isolated environments where AI agents can write and execute code safely. The repo contains Terraform configs and Go services for orchestrating microVMs on GCP and AWS, letting LLMs run arbitrary code without the arbitrary damage.

The interesting bit

Most “AI infrastructure” repos are thin wrappers around Kubernetes with a LangChain import. E2B’s stack is lower-level: Firecracker microVMs orchestrated through Nomad and Consul, which means actual sandboxing via KVM rather than container escape-hatch roulette. They’re betting that agent workloads need real VMs, not just namespaced processes.

Key highlights

• Deploys on GCP (stable) and AWS (beta); Azure and bare metal are still unchecked boxes on the roadmap
• Built in Go with Terraform for orchestration, plus HashiCorp’s Nomad/Consul stack for scheduling and service discovery
• Designed for self-hosting if you don’t want E2B’s managed cloud
• Sibling to the main e2b-dev/e2b repo, which holds the SDKs and CLI that developers actually touch

Caveats

• AWS support is explicitly marked beta; Azure is a TODO
• The README is essentially a signpost—most detail lives in self-host.md and CONTRIBUTING.md, which aren’t excerpted here

Verdict

Worth studying if you’re building multi-tenant code execution for LLMs and need to understand what “production-grade sandboxing” actually costs in complexity. Skip it if you just want to call an API and not think about Nomad job specs.