Docker Sandboxes vs E2B vs Cloudflare Sandboxes: which one fits your AI agent?

Most comparisons of Docker Sandboxes, E2B, and Cloudflare Sandbox SDK treat them as three interchangeable hosted APIs, differing only on price and latency. They aren't interchangeable. Docker Sandboxes is a local developer-workstation tool. The sbx CLI runs on your own machine, mounts your real filesystem, and spins up a microVM on your own hardware. E2B and Cloudflare Sandbox SDK are both cloud-hosted, API-driven, billed remote execution services built for running other people's agent sessions at scale. Putting all three in one flat feature table, the way most existing comparisons do, obscures that split.

The real question isn't "which is fastest or cheapest." It's which shape of sandbox your situation needs, a local isolation layer around your own coding agent, or a hosted execution backend for a product other people use. This article answers that question directly, with a verified feature comparison, real pricing math, and a working demo of each product's setup.

The quick answer

What each one is

Feature comparisons usually start with isolation primitives and pricing tiers. Before any of that, it's worth being plain about what each of these products is, because that's where the existing comparisons blur together three things that aren't the same shape.

Is Docker Sandboxes a cloud service or local-only?

Local-only. There's no Docker-hosted "call an API, get a remote sandbox" product. The sbx CLI installs locally. On macOS that's brew trust docker/tap and brew install docker/tap/sbx, on Windows it's winget install -h Docker.sbx, and on Ubuntu it's curl -fsSL https://get.docker.com | sudo REPO_ONLY=1 sh followed by sudo apt-get install docker-sbx. After sbx login, running sbx run claude from a project directory launches an agent inside a sandbox on your own machine.

That sandbox mounts your actual working tree. Docker's architecture docs describe it directly. "Your workspace is mounted directly into the sandbox through a filesystem passthrough. The sandbox sees your actual host files, so changes in either direction are instant with no sync process involved." Outbound network traffic routes through a proxy on the host itself, not through any Docker-operated cloud infrastructure. Docker's own comparison to alternatives sums up the tradeoff. "Sandboxes trade higher resource overhead (a VM plus its own daemon) for complete isolation. Use containers when you need lightweight packaging without Docker access. Use sandboxes when you need to give something autonomous full Docker capabilities without trusting it with your host environment."

What is E2B and how does it isolate code?

E2B is a hosted API and SDK (Python, TypeScript) for spinning up cloud-run sandboxes. Per E2B's own docs, "E2B provides isolated sandboxes that let agents safely execute code, process data, and run tools. Our SDKs make it easy to start and manage these environments." Each sandbox runs as a Firecracker microVM, the same virtualization technology AWS Lambda uses, with its own kernel. E2B's homepage describes it plainly. "Each sandbox is powered by Firecracker, a microVM made to run untrusted workflows."

E2B is agent and framework agnostic rather than tied to a specific coding agent. Its pricing page lists named customers building on it, including Rogo, Hugging Face, Manus, Groq, Lindy, and Gumloop, spanning use cases from AI-generated code execution to replicating research results to powering agent workflows.

Does Cloudflare Sandbox SDK use microVMs?

No, and this is worth answering precisely rather than flattening it. Cloudflare Sandbox SDK is container-based. It's built directly on Cloudflare Containers and deployed as part of your own Worker, with each sandbox backed by a dedicated Durable Object. Cloudflare's own docs describe it this way. "Built on Containers, Sandbox SDK provides a simple API for executing commands, managing files, running background processes, and exposing services, all from your Workers applications... Each sandbox runs in its own isolated container with a full Linux environment, providing strong security boundaries while maintaining performance."

Cloudflare's platform does add a VM layer underneath that container, but it's a different and less consumer-visible model than E2B's or Docker's per-sandbox microVM. Cloudflare's container architecture docs state it plainly. "Each container instance runs inside its own VM, which provides strong isolation from other workloads running on Cloudflare's network." That's a platform-level tenant-isolation detail, not a per-sandbox microVM boundary you interact with directly the way you do with E2B or Docker Sandboxes. It's also a separate thing entirely from Cloudflare's Dynamic Worker Loader, a V8-isolate-based product that has nothing to do with Sandbox SDK.

Feature and architecture comparison

Two rows deserve more than a table cell. Docker Sandboxes ships a built-in MCP gateway, a single endpoint that supported agents connect to, which brokers access to registered MCP servers separately from the network proxy. Neither E2B nor Cloudflare Sandbox SDK documents an equivalent first-class MCP integration. Docker also has Docker AI Governance, an org-wide policy layer for network, filesystem, and MCP controls enforced uniformly across every developer's machine, something neither hosted competitor has a documented counterpart for. Most existing comparisons skip both of these entirely.

Pricing compared, with a real cost example

Docker Sandboxes itself costs nothing. The sbx CLI is free for commercial use, and the only paid layer is Docker AI Governance, the org-wide policy add-on, which has no public list price and requires talking to Docker's sales team. There's no marginal cloud compute cost either, since everything runs on your own hardware.

E2B's Hobby tier is free, with a one-time $100 of usage credits, no credit card required, sandbox sessions up to 1 hour, and up to 20 concurrently running sandboxes. The Pro tier is $150/month plus usage, with sessions up to 24 hours and up to 100 concurrent sandboxes (extendable to 1,100). Usage is billed per second. At the default 2 vCPU size, compute costs $0.000028 per second, with memory billed separately at $0.0000045 per GiB-second.

Cloudflare Sandbox SDK has no standalone free tier. It requires the Workers Paid plan at $5/month minimum before any usage. On top of that, containers are billed per 10ms of active runtime, at $0.000020 per additional vCPU-second beyond 375 included vCPU-minutes per month, and $0.0000025 per additional GiB-second beyond 25 included GiB-hours per month. Cloudflare also bills separately for the Worker and the Durable Object backing each sandbox.

A worked example makes the difference concrete. Take 1,000 five-minute agent sessions at a 2 vCPU sandbox size, a realistic load for a small agent product.

That last row is the actual punchline. For this exact workload, Docker Sandboxes has no cloud billing at all, because there's no cloud compute involved in the first place. The comparison only makes sense once you accept that Docker Sandboxes isn't competing on per-second pricing, it's not paying for compute the same way.

Try it

Try Docker Sandboxes locally

Docker Sandboxes needs a local sbx login and, on Linux, adding your user to the kvm group for hypervisor access. Here's the install and run sequence for each platform.

# macOS
brew trust docker/tap
brew install docker/tap/sbx
sbx login

# Windows
winget install -h Docker.sbx
sbx login

# Linux (Ubuntu)
curl -fsSL https://get.docker.com | sudo REPO_ONLY=1 sh
sudo apt-get install docker-sbx
sudo usermod -aG kvm $USER
newgrp kvm
sbx login

Once you're logged in, launch an agent in a sandbox from your project directory:

cd ~/my-project
sbx run claude

Try E2B in a few lines of Python

E2B's Hobby tier is free, with $100 in usage credits and no credit card required. Install the SDK, get an API key from the E2B dashboard, and run code inside an isolated sandbox in a handful of lines.

pip install e2b-code-interpreter

from e2b_code_interpreter import Sandbox

with Sandbox.create() as sandbox:
    sandbox.run_code("x = 1")
    execution = sandbox.run_code("x += 1; x")
    print(execution.text)  # outputs 2

Running this against a real E2B sandbox prints 2, confirming the sandbox executed both statements and persisted state between the two run_code calls.

Try Cloudflare Sandbox SDK in a Worker

Cloudflare Sandbox SDK is currently mid-transition to a 1.0 release, available as a preview on the npm @next tag. Cloudflare recommends new projects start there rather than on the current stable package. Scaffold a project from the minimal example template:

npm create cloudflare@latest -- my-sandbox --template=cloudflare/sandbox-sdk/examples/minimal
cd my-sandbox

import { getSandbox, proxyToSandbox, type Sandbox } from "@cloudflare/sandbox";

export { Sandbox } from "@cloudflare/sandbox";

type Env = {
  Sandbox: DurableObjectNamespace<Sandbox>;
};

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    const url = new URL(request.url);
    const sandbox = getSandbox(env.Sandbox, "my-sandbox");

    if (url.pathname === "/run") {
      const result = await sandbox.exec('python3 -c "print(2 + 2)"');
      return Response.json({
        output: result.stdout,
        error: result.stderr,
        exitCode: result.exitCode,
        success: result.success,
      });
    }

    if (url.pathname === "/file") {
      await sandbox.writeFile("/workspace/hello.txt", "Hello, Sandbox!");
      const file = await sandbox.readFile("/workspace/hello.txt");
      return Response.json({
        content: file.content,
      });
    }

    return new Response("Try /run or /file");
  },
};

Deploy with npx wrangler deploy. This requires a Workers Paid plan ($5/month minimum), since Sandbox SDK has no standalone free tier.

Which one should you use

The decision comes down to where your agent runs and who it runs for, not raw feature counts.

• Solo developer running Claude Code, Copilot CLI, Codex, or OpenCode on your own laptop, wanting isolation without giving up local filesystem access → Docker Sandboxes. Docker's product page lists Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro as supported out of the box, plus the option to add your own.
• Building a SaaS or agent platform that spins up isolated execution per end-user request at scale, needing hardware-level microVM isolation → E2B. Its Pro tier is explicitly built for this, supporting up to 100 concurrent sandboxes with the option to purchase extra concurrency up to 1,100.
• Already all-in on Cloudflare Workers and Durable Objects, comfortable with container-level isolation, wanting sandboxed execution co-located with your edge app → Cloudflare Sandbox SDK. It's deployed as part of a Worker and requires a Durable Objects binding, so it's most natural if you're already building there.

Skip Docker Sandboxes if you need to run sandboxes for other people's requests rather than your own local agent session, need hardware-level isolation guarantees rather than a dev-loop tool, or don't want to manage local hypervisor access (kvm group membership on Linux).

Skip E2B if you're not building a hosted product that needs remote sandbox execution, you're already committed to the Cloudflare stack, or you need session lengths beyond the 24-hour Pro-tier ceiling without relying on pause/resume.

Skip Cloudflare Sandbox SDK if you want a free tier to experiment with first, you're not already using Workers or Durable Objects, or you need per-sandbox microVM isolation rather than container-level isolation.

FAQ

Is Docker Sandboxes a cloud service or local-only?

Local-only. It runs via the sbx CLI on your own machine, and there's no hosted-API equivalent documented anywhere in Docker's own docs.

Does Cloudflare Sandbox SDK use microVMs?

No. It's container-based, though Cloudflare's platform additionally runs each container inside its own VM for tenant isolation, a different model from E2B's or Docker's per-sandbox Firecracker microVM.

Which AI agent sandbox is cheapest?

It depends on your situation rather than one flat answer. Docker Sandboxes has $0 marginal cost if you already have local hardware and don't need the paid governance layer. E2B has a free Hobby tier with no credit card, as long as you stay within the $100 in credits. Cloudflare Sandbox SDK has no free tier at all, requiring the $5/month Workers Paid plan minimum before any sandbox usage.

What's the best sandbox for AI coding agents?

There's no single winner, it depends on the shape of the problem. If you're running your own coding agent locally, Docker Sandboxes fits. If you're building a hosted product for other people's agent sessions at scale, E2B fits. If you're already on Cloudflare Workers, Cloudflare Sandbox SDK fits.

Can I self-host any of these?

Docker Sandboxes is inherently local by design, so there's no separate self-hosting question, it always runs on your own machine. E2B is confirmed to support self-hosted and BYOC deployments in your own AWS, GCP, or Azure account or VPC, per E2B's own site and changelog. Cloudflare Sandbox SDK has no documented self-host option, it's a Cloudflare-platform-only product built on Cloudflare Containers.