Purpose
E2B is positioned as the "AI Agent Cloud," that is, providing an isolated machine for each AI agent. Users are responsible for delivering the agent (model, prompts, agent logic), and E2B is responsible for running the infrastructure beneath it. It is suitable for workloads such as coding, computer use, background jobs, and reinforcement learning rollouts, providing an isolated microVM for each session through a unified API.
Main capabilities
Each session runs a complete Linux microVM, and API capabilities are provided around it:
- Execute code: supports running Python or JavaScript; Code Interpreter builds on this to provide
runCode(), and can return charts, tables, and images. - Computer use: Desktop / Use Computer provides a screen, mouse, keyboard, and live stream on top of the sandbox, for computer-use agents to drive desktop applications.
- Session control: can pause, resume, snapshot, and fork (
sandbox.pause()). - Networking: supports inbound and outbound, and follows user policies (
sandbox.getHost(port)); under outbound policies, the sandbox accesses the outside through the policy rather than bypassing it. - Storage: provides a persistent file system, and can connect to the user's own storage bucket (
sandbox.files.write()). - Observability: provides metrics, logs, and signed webhooks (
sandbox.getMetrics()). - Secret management: environment variables are isolated per sandbox, and secrets are available to the agent but cannot be read.
- Tool calling: can call 200+ MCP servers from inside the sandbox; supports Anthropic, OpenAI, and open-weight models.
Officially, each sandbox runs on a proprietary MicroVM runtime based on Firecracker, with an independent kernel, as well as custom memory and snapshot layers; sandboxes start from snapshots, so each session gets its own kernel at container speed. The platform has been certified for SOC 2 Type II and HIPAA compliance.
How to use
Users can use E2B through the SDK, CLI, and API. The official installation command given is pip install e2b or npm i e2b, and offers a free tier. The platform supports working with existing agent stacks; the officially listed SDKs and frameworks include Claude Managed Agents SDK, OpenAI Agents SDK, LangChain Deep Agents + Open SWE, Vercel AI SDK + Eve; it also provides prebuilt templates (such as OpenCode) and supports creating custom templates. The runtime is open source under Apache-2.0, and related projects are available on GitHub (such as e2b-dev/E2B, fragments, code-interpreter, desktop, runtime, surf).
Deployment and Use Cases
Sandboxes can run on E2B Cloud (US, EU, Asia-Pacific regions) or in the user's own AWS, GCP, or Azure account; the official description states that the same SDK, CLI, and API are universal across all environments, and migration only requires changing configuration.
The officially listed typical scenarios include:
- Turning prompts into runnable applications: a prompt starts a coding sub-agent, builds the application, and then E2B serves it in real time.
- Agents that keep working after the notebook is closed: a sandbox continuously uses the browser, CLI, and storage, and scales compute on its own.
- Analytics that compliance teams can approve: connect to multiple sources, run analysis, and return a complete report.
- Isolated per customer, running in their own cloud: each customer's agent runs in its own microVM, deployed via BYOC in the user's AWS, GCP, or Azure account.
- Sensitive data processing: genomic, compound, and clinical data flow into a secure compute environment, reports are output, and data does not leak out.
- Agents that use applications like a human: a full desktop with screen, mouse, and keyboard; the agent drives desktop applications and streams in real time.
- Reinforcement learning: quickly launch, score, and clean up thousands of machines.
- World model agent simulation: run game simulations to generate synthetic training data.
- Agent evaluation: score outputs in parallel, then assign scores.
- Agent benchmarking: use the same prompt for each agent on the same machine and score them side by side.
Deployment Options and Limits
The officially listed deployment options include: E2B Cloud (managed by E2B, billed by the second, with a free tier, covering US, EU, and Asia-Pacific regions, SOC 2 Type II, HIPAA); BYOC (deployed in the user's AWS, GCP, or Azure account, with sandboxes and data remaining within the user's VPC, using the user's IAM, KMS, and audit logs); Private Cloud (the entire platform is within the user's boundary, the control plane is in the user's account, aimed at air-gapped and sovereign networks, and is in development); E2B Embed (deploys the full E2B stack on a single node, can be delivered with a product or deployed in a customer tenant, supports Docker Compose, Terraform on AWS or GCP, or Kubernetes, open source Apache-2.0).
Free tier limits: get $100 one-time usage credit, sessions up to 1 hour, up to 20 concurrent sandboxes, use default sandbox CPU and RAM, no credit card required. Paid tiers are billed by the second, sessions up to 24 hours, 100 concurrent sandboxes (scalable to 1,100), support custom CPU and RAM, and are billed through Stripe. Higher tiers provide committed usage, custom billing, multi-day persistent sessions, tens of thousands of concurrent sandboxes, custom specifications, and BYOC deployment, with customized deployment and billing.
The above information comes from official public descriptions; other features, prices, or limits not listed here are not currently provided.
