e2b-dev/runtime

The runtime behind every E2B stack: Cloud, Enterprise, and your own machine.

What it solves

E2B Runtime provides a secure, isolated environment for AI agents to execute untrusted code. It eliminates the slow boot times of traditional virtual machines by using snapshot-based restoration, allowing agents to have their own dedicated Linux machines that can be paused, resumed, and forked instantly.

How it works

The system leverages Firecracker microVMs and a custom architecture to separate the control plane (API) from the data plane (execution). Instead of booting a kernel from scratch, it restores pre-booted VM templates from object storage. It uses userfaultfd for lazy memory loading and a copy-on-write overlay for the filesystem, ensuring only necessary data is fetched. Inside each VM, a specialized agent called envd provides a REST and Connect RPC API for managing processes, files, and ports.

Who it’s for

  • AI Agent Developers: Teams needing precise control over the sandbox environment and local-to-production parity.
  • Platform Teams: Organizations requiring agent execution within their own cloud infrastructure to avoid "black box" solutions.
  • Infrastructure Engineers: Those interested in high-scale snapshot-based scheduling, Firecracker microVMs, and lazy memory restoration.

Highlights

  • Hardware-isolated sandboxes: Each session runs in its own Firecracker microVM with dedicated network namespaces and egress firewalls.
  • Instant Lifecycle: Ability to pause, resume, and fork running sandboxes via memory and disk diffing.
  • Template System: Layered builds from Docker images with cached layers and boot-optimization passes.
  • Integrated API: envd allows SDKs to interact with the VM's filesystem and processes without dropping workloads during upgrades.
  • Dynamic Routing: Automatic sandbox URLs for any process that opens a port, routed via an edge proxy.

Related

  • Project
    m0n0x41d/haft

    A local governance layer for AI coding agents that provides durable, structured project memory and enforces rigorous engineering reasoning via the First Principles Framework.

  • Project
    turboderp-org/exllamav3

    An inference library for running local LLMs on consumer GPUs featuring EXL3 quantization, parallel inference, and CPU offloading for large MoE models.

  • Project
    openclaw/crabbox

    A remote software testing and execution control plane that allows developers and AI agents to run repository commands on managed remote runners instead of local machines.

  • Project
    Minglink/dsh-infinite-gen-4

    A cybersecurity red-teaming and robustness benchmarking suite for DeepSeek-V4 series models, using dual-layer prompt injection to test model defenses.