E2B
Firecracker microVM sandboxes for running AI-generated code. Python and JS SDKs, 24-hour sessions, pause and resume with full memory state, Apache 2.0 core you can self-host.
Overview
Sandboxes for AI-generated code, built on a fork of Firecracker, the same microVM technology under AWS Lambda. You start one with an SDK call, it boots in well under a second, and the agent inside gets a full Linux userspace with filesystem, network and process control. Sessions run up to 24 hours on Pro. The company reports over a billion sandboxes started, 3.5M monthly SDK downloads and use across most of the Fortune 100, with Hugging Face, Perplexity, Groq and Manus among named customers.
The product line has widened past plain code execution. The Code Interpreter SDK handles the stateful REPL case, Desktop Sandbox adds a virtual display for computer-use agents, persistent volumes survive sandbox death, and template builds bake dependencies from a Dockerfile. Sandbox persistence, in public beta since early 2025, pauses a sandbox to storage with memory intact and resumes it later or forks copies from the snapshot, which is what makes long-lived agent sessions affordable instead of pinning a VM for hours.
Licensing is unusually clean for this category. The SDKs, the code-interpreter and desktop layers and the production infra repo are all Apache 2.0 and self-hostable on AWS, GCP, Azure or bare Linux. E2B raised a $21M Series A led by Insight Partners in July 2025, taking total funding to $32M, and remains independent. Competition thickened through 2026: Daytona, Cloudflare Sandboxes, Vercel Sandbox and Modal all shipped comparable primitives, and Modal wins outright when the sandbox itself needs a GPU.
Key Features
- ✓ Firecracker microVM isolation per sandbox, a kernel boundary rather than a shared container runtime
- ✓ Python and JavaScript SDKs plus a REST Sandbox API and the e2b CLI, with first-party support in the OpenAI Agents SDK
- ✓ Sandbox persistence: pause and resume restores filesystem and full memory state, and snapshots fork new sandboxes from a saved state
- ✓ Desktop Sandbox with a virtual display for computer-use agents that need a mouse and a screen
- ✓ Custom sandbox templates built from Dockerfiles plus persistent volumes, so cold starts do not pay for pip install
- ✓ Self-hostable on AWS, GCP, Azure or bare Linux from the Apache 2.0 infra and forked-Firecracker repos
Where it holds
- • The isolation model is the real thing: microVMs, so untrusted model output gets a hardware boundary instead of a namespace
- • The production infra is Apache 2.0, which makes an on-prem or air-gapped deployment a config problem rather than a vendor negotiation
- • Wide adoption at Hugging Face, Perplexity, Groq and Manus means the sharp edges are documented by other people first
- • Per-second billing with sub-second starts fits bursty agent traffic far better than keeping a pool of VMs warm
Where it breaks
- • Managed cloud runs CPU-only hosts. A sandbox that needs a GPU means self-hosting the open-source stack on your own bare metal.
- • The $150/month Pro floor is steep for a side project, and the free tier's $100 credit is one-time rather than monthly
- • 1,100 concurrent sandboxes is the ceiling on self-serve; past that you are in enterprise sales
- • Self-hosting is real work. Terraform, Nomad and a forked Firecracker, not a docker compose up.
My Take
Running model-written code on your own infrastructure is a bad idea, and this is the least painful way to stop doing it. Firecracker microVMs give a real kernel boundary, pause and resume now restores full memory state rather than just the filesystem, and the whole infra repo is Apache 2.0 if compliance says the code cannot leave your VPC. The catch is GPUs. Managed hosts are CPU-only, so a sandbox that needs to run a vision model locally sends you to self-hosted bare metal or to Modal instead.
Quick Info
- Pricing:
- freemium
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Hobby is free with a one-time $100 usage credit, 20 concurrent sandboxes, a 1-hour session cap and 10 GiB storage. Pro is $150/month for 100 concurrent sandboxes, 24-hour sessions, configurable CPU and RAM and 20 GiB storage, with extra concurrency purchasable up to 1,100. Compute meters per second on top: $0.000014/s per vCPU ($0.000028/s for the default 2 vCPU) and $0.0000045/GiB/s of RAM between 512 MiB and 8,192 MiB. Enterprise is quoted. Self-hosting the Apache 2.0 stack costs only your own cloud bill.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where E2B holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to E2B
Modal
freemiumServerless Python compute with per-second GPU billing, sub-second cold starts and gVisor Sandboxes for agent code. H100 at $0.001097/s, scale to zero when idle.
AgentScope
open-sourceApache 2.0 agent framework from Alibaba Tongyi Lab's SysML team. Version 2.0 folded the separate Runtime project into the core, so sandboxing, agent serving and observability now ship as one...