The agent cloud

Every agent gets a machine

E2B runs the isolated machine your agent works on. You ship the agent. We run the infrastructure. Our cloud or yours.

Start building Talk to an engineer

i4h2k9x@ us-east-1us-east-1

$ e2b sandbox create

✓ ready i4h2k9x · from snapshot

agent> plan: build checkout

agent> write app/page.tsx +214

agent> write checkout.ts +86

agent> bun test

✓ 14 passed — 2.1s

agent> bun run build

✓ build ok — 1.9s

agent> deploy i4h2k9x.e2b.app

i4h2k9x.e2b.app

Checkout

$48

$32

$48

Total $128

pay now

booted from snapshotuptime 04:37:12

100K+ teams on E2B · 94 of the Fortune 100 signed up · top AI labs & the fastest AI startups run here

Where E2B fits

Your agent is the product.
E2B is where it runs.

You bring the model, prompts, and agent logic. E2B runs the machine underneath. Coding, computer use, background jobs, RL rollouts: one API, an isolated microVM per session, in our cloud or your AWS or GCP account.

Your agents

Any agent, any harness.

OpenAI Agents SDK · Claude Code · Codex · your own harness

Any agentcoding · data · computer use · evals · RL

E2B · the boundary

A full machine per agent session.

a microVM per session · Apache-2.0 runtime · SOC 2 Type II

The sandboxboots from a snapshot · shell · filesystem · browser · network
Lifecyclesandbox.pause()
Networksandbox.getHost(port)
Storagesandbox.files.write()
Observabilitysandbox.getMetrics()

Through the boundary

Any model, any tool, your data, your keys.

the sandbox reaches out under your egress policy, never around it

Model callsAnthropic · OpenAI · open weights
MCP tools200+ servers, called from inside the sandbox
Your databuckets · repos · internal APIs
Secretsenv vars scoped per sandbox

The cloud beneath

Our cloud, or inside yours.

same SDK, same API, wherever it runs

E2B CloudUS · EU · APAC
Your cloudAWS · GCP

Use cases

Build and train
any agent.

Agents that act, use tools, and write and run code in optimized environments

Turn prompts into running apps
A prompt spins up coding sub-agents, builds the app, then serves it live from E2B.
Case study →
Agents that work while the laptop is closed
One sandbox keeps using browser, CLI, and storage, scaling compute on its own.
Analysis your compliance team can approve
Ingest many sources, run the analysis, return a full report.
Every customer isolated, in your own cloud
Each customer's agent runs in its own microVM, inside your AWS or GCP account via BYOC.
See E2B for enterprise →
Sensitive data in, report out. Nothing leaves.
Genomes, compounds, and clinical data flow into secure compute; a report comes out.
Agents that use apps like a person
A full desktop with screen, mouse, and keyboard. The agent drives desktop apps, streamed live.
Case study →
Turn prompts into running apps
A prompt spins up coding sub-agents, builds the app, then serves it live from E2B.
Case study →
Agents that work while the laptop is closed
One sandbox keeps using browser, CLI, and storage, scaling compute on its own.
Analysis your compliance team can approve
Ingest many sources, run the analysis, return a full report.
Every customer isolated, in your own cloud
Each customer's agent runs in its own microVM, inside your AWS or GCP account via BYOC.
See E2B for enterprise →
Sensitive data in, report out. Nothing leaves.
Genomes, compounds, and clinical data flow into secure compute; a report comes out.
Agents that use apps like a person
A full desktop with screen, mouse, and keyboard. The agent drives desktop apps, streamed live.
Case study →
Turn prompts into running apps
A prompt spins up coding sub-agents, builds the app, then serves it live from E2B.
Case study →
Agents that work while the laptop is closed
One sandbox keeps using browser, CLI, and storage, scaling compute on its own.
Analysis your compliance team can approve
Ingest many sources, run the analysis, return a full report.
Every customer isolated, in your own cloud
Each customer's agent runs in its own microVM, inside your AWS or GCP account via BYOC.
See E2B for enterprise →
Sensitive data in, report out. Nothing leaves.
Genomes, compounds, and clinical data flow into secure compute; a report comes out.
Agents that use apps like a person
A full desktop with screen, mouse, and keyboard. The agent drives desktop apps, streamed live.
Case study →

Reinforcement learning
Thousands of machines spin up, grade, and clear, fast.
Labs under NDA
World model agentic simulations
Run game sims to generate synthetic training data.
Labs under NDA
Agent Evaluations
Outputs graded in parallel, then scored.
Agent benchmarks
Same prompt for every agent, in the same machine, graded side by side.
Reinforcement learning
Thousands of machines spin up, grade, and clear, fast.
Labs under NDA
World model agentic simulations
Run game sims to generate synthetic training data.
Labs under NDA
Agent Evaluations
Outputs graded in parallel, then scored.
Agent benchmarks
Same prompt for every agent, in the same machine, graded side by side.
Reinforcement learning
Thousands of machines spin up, grade, and clear, fast.
Labs under NDA
World model agentic simulations
Run game sims to generate synthetic training data.
Labs under NDA
Agent Evaluations
Outputs graded in parallel, then scored.
Agent benchmarks
Same prompt for every agent, in the same machine, graded side by side.

The machine

Everything
your agent needs

One isolated microVM per session, with an API around it. Run code, control the network, keep files, hold secrets, watch it, pause or fork it.

Execution

Commands, code, a PTY, a full desktop.

Learn more →

Network

Ingress and egress, under your policy.

Learn more →

Storage

A filesystem that persists, plus your buckets.

Learn more →

Observability

Metrics, logs, and signed webhooks.

Learn more →

Secrets vault

Keys your agent can use but never read.

Learn more →

Lifecycle

Pause, resume, snapshot, fork.

Learn more →

Sandbox

A full Linux microVM per agent session. Every capability above is this one machine.

npm: e2b
PyPI: e2b

Code Interpreter

The sandbox plus runCode(): execute Python or JavaScript, get charts, tables, and images back.

npm: @e2b/code-interpreter
PyPI: e2b-code-interpreter

Desktop · Use Computer

The sandbox plus a screen: mouse, keyboard, and a live stream for computer-use agents.

npm: @e2b/desktop
PyPI: e2b-desktop

Deployment

Our cloud or yours.
Same API. One standard.

Sandboxes run where your data is allowed to live. Same SDK, CLI, and API everywhere. Moving is a config change.

Managed by E2B. Nothing to host.

Billed by the second, free tier to start
US, EU, and APAC regions
SOC 2 Type II, HIPAA

E2B CloudControl plane + SandboxesYou connect via the SDK. Nothing to host.

The runtime is Apache-2.0. Self-host it with Terraform today, no contract needed.

See E2B for enterprise

Security

Full isolation.
Container speed.

Every sandbox runs on E2B's proprietary MicroVM runtime, built on Firecracker with dedicated kernels and a custom memory and snapshot layer. It safely runs untrusted agentic workloads across tens of thousands of parallel sessions.

Battle-tested across 1B+ sandboxes, MicroVMs come standard. E2B is SOC 2 Type II and HIPAA compliant.

Visit the trust center

VMs are slow to start. Ours boot from a snapshot, so every session gets its own kernel at container speed.

► E2B by the numbers

94Of the Fortune 100 signed up

100K+Teams on E2B

1B+Sandboxes started

10M+Monthly SDK downloads

Customers

In their words

“E2B allows us to scale-out training runs by launching hundreds of sandboxes in our experiments, which was essential in Open R1.”

— Lewis Tunstall, Research Engineer

Read the case study

“Manus doesn't just run some pieces of code. It uses 27 different tools, and it needs E2B to have a full virtual computer to work as a real human.”

— Tao Zhang, Co-founder

Read the case study

“E2B gives us a fast, secure way to run AI-generated code at scale, so we can ship workflow automation our enterprise customers trust.”

— Rahul Behal, Co-founder

Read the case study

“We had implemented E2B in a week, needing just a single engineer working on that in spare cycles. Building this in-house would take us a few weeks and require multiple people.”

— Luiz Scheidegger, Head of Engineering

Read the case study

Works with any agent stack.

Start with a ready-to-run template, or bring the SDKs and frameworks you already use. E2B provides the isolated machine underneath.

Pre-built templates

Create your own

SDKs and frameworks

CrewAI
Agno
Mastra
Claude Managed Agents SDK
OpenAI Agents SDK
Google ADK
LangChain Deep Agents + Open SWE
Crabbox
Vercel AI SDK + Eve
Flue

open source

Open at the core

The runtime your agents run on is Apache-2.0: readable, self-hostable, and the base of some of the most-used projects in the OSS agent ecosystem. The only open-source hosted microVM sandbox platform.