E2B

E2B provides secure cloud sandboxes for running AI-generated code. We built this as open-source infrastructure to solve a critical problem: AI agents need safe, isolated environments to execute untrusted code without compromising host systems.

Our SDK lets you create sandboxes in seconds, execute code, interact with filesystems, and manage processes—all through simple APIs in Python or JavaScript. See GitHub repository.

🏗️ System Architecture

We use a three-layer architecture: Client SDKs communicate with our REST API to manage sandboxes, then connect directly via gRPC to the envd daemon running inside each microVM. This hybrid approach (REST for management, gRPC for operations) reduces latency for filesystem/process operations. See SDK source.

📦 Repository Structure

🔧 packages/cli/ - Template management CLI (TypeScript, Commander.js). See CLI source
📘 packages/js-sdk/ - JavaScript/TypeScript SDK (openapi-fetch + Connect RPC). See JS SDK
🐍 packages/python-sdk/ - Python SDK (httpx + protobuf). See Python SDK
🌐 apps/web/ - Next.js app (dashboard + docs + SDK reference). See Web app

💡 Why This Architecture

We chose Firecracker microVMs (not Docker containers) for sandboxes because they provide kernel-level isolation—each sandbox has its own kernel, preventing escape attacks. The trade-off: 150ms boot time vs. 50ms for containers, but worth it for security. See infrastructure repo.

We use gRPC for operations (filesystem, process) because it's 3-5× faster than REST for streaming stdout/stderr. REST handles sandbox lifecycle (create, kill) because it's simpler for one-off operations. See RPC client.