AI Tools

What Are Docker Sandboxes for AI Agents?

Published 2026-08-11 · Researched and price-checked on official sources

Docker Sandboxes are free, disposable microVM environments that let AI coding agents like Claude Code, Copilot CLI, and Codex run unattended without touching your host machine. Docker launched them on August 10, 2026. The sandbox CLI is free to use; Docker Business, the paid tier for team security and compliance, costs $24 per user per month.

Key facts

What matters

Docker Sandboxes are free local microVM environments for AI coding agents, announced August 10, 2026 on Hacker News (626 points).
Out-of-the-box support for Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro, plus custom agents.
The sbx CLI runs on macOS, Windows, and Ubuntu, and Docker Desktop is not required.
Sandboxes are free to start; Docker Business costs $24 per user per month, and AI Governance is priced through sales.
Paid cloud alternatives: E2B starts with $100 in credits (Pro at $150/month) and Daytona includes $200 in free compute.

What are Docker Sandboxes?

Docker Sandboxes is a new Docker tool for running AI coding agents safely. Each agent runs inside a disposable microVM on your own machine, with your dev environment and project workspace mounted in and the host filesystem left untouched. Docker announced it on Hacker News on August 10, 2026, and the thread passed 620 points within a day.

It targets the way modern coding agents work: tools like Claude Code, Copilot CLI, and Codex run unattended, install packages, edit configs, and execute commands without a human approving each step. Docker's answer is a hard boundary around that autonomy. Sandboxes are a standalone CLI called sbx, not a Docker Desktop feature.

How Docker Sandboxes work

Docker's product page FAQ describes a sandbox as "a microVM isolated environment that protects your filesystem and network from agents running inside it." Each agent gets a dedicated microVM with only the project workspace mounted, so installs, config edits, and service restarts stay inside the boundary.

Unattended execution: no permission prompts and no manual review.
Agents can install packages, modify configs, and spin up their own Docker containers inside the sandbox.
Disposable by default: one command tears the environment down.
Permissive agent flags are the supported default: the page pairs YOLO mode (--dangerously-skip-permissions) with microVM isolation and calls that combination safe.

Network and filesystem controls are configurable per sandbox, and Docker AI Governance extends those policies org-wide.

Which coding agents does Docker Sandboxes support?

Docker lists Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro as supported out of the box, and custom agent harnesses can be added. The demo on the product page shows an agent installing packages, running Docker, and modifying configs, then disposing of the sandbox in one command.

Installation is a package manager command: brew install docker/tap/sbx on macOS, winget install Docker.sbx on Windows, and apt-get install docker-sbx on Ubuntu after Docker's repo setup script. Docker's FAQ confirms sandboxes do not require Docker Desktop.

The current friction point is a Docker account login for the CLI, which commenters on the launch thread flagged as a hurdle for a local-only tool.

Docker Sandboxes pricing: free CLI, paid governance

Docker Sandboxes is free to get started: the product page's primary call to action is "Get started free." Docker's pricing page lists four plans from Personal at $0 to Business at $24 per user per month, and Business is the tier aimed at organizations that need security and compliance controls.

Team-wide enforcement is where the paid layer starts. Docker AI Governance adds network access policies, filesystem rules, and org-wide MCP governance, but it has no public price: inquiries go through a sales form. If you run agents only on your own machine, the free tool covers the core workflow.

Docker Sandboxes vs E2B vs Daytona

The key difference is where sandboxes run. Docker Sandboxes is local: microVMs on your machine, no per-hour compute bill, but no cloud concurrency or managed fleet either. E2B and Daytona are cloud sandbox services with usage-based pricing.

E2B bills per use. New users get $100 in one-time credits, Hobby is $0 per month with a one-hour max continuous runtime, and Pro is $150 per month for 24-hour runtimes and up to 1,100 concurrent sandboxes, per its billing documentation. E2B publishes integration guides for Claude Code and Codex.

Daytona offers pay-as-you-go cloud sandboxes with $200 in included compute. Rates are per hour: vCPU at $0.0504, Nvidia H100 at $3.95, H200 at $4.54, RTX 4090 at $0.99, memory at $0.0162 per GiB, and storage at $0.000108 per GiB after 5 free GiB, per Daytona's pricing page. GPU instances are Daytona's edge over local sandboxes.

Who should use Docker Sandboxes

Use it if you run a CLI coding agent on your own machine and want unattended execution without host access. That is the exact case Docker built for: local microVM isolation, no per-hour bill, disposable environments.

Skip it if agents must run in CI, on servers, or across a team. Those workloads need cloud sandboxes: E2B for API-driven fleets, Daytona when GPU instances matter. Plan for $24 per user per month only if your organization needs Docker Business-level controls, and treat AI Governance as a sales-priced enterprise add-on.

Current caveats: sandboxes run on your hardware, the CLI needs a Docker login, and the product launched this week, so expect integrations to evolve. The supported agent list and pricing above are verified from Docker's pages as of August 11, 2026.

At a glance

	Docker Sandboxes	E2B	Daytona
Pricing model	Free local CLI	Usage-based, Hobby $0/month	Pay-as-you-go
Free credits	Free to start	$100 one-time	$200 included
Where sandboxes run	Local microVMs	Cloud	Cloud
Max continuous runtime	Until you dispose it	1h on Hobby, 24h on Pro	Per-hour billing
Best for	Local-first solo developers	Teams scaling cloud sandboxes	GPU-heavy agent workloads

Docker Sandboxes vs E2B vs Daytona, verified August 11, 2026

FAQ

Are Docker Sandboxes free?

Yes, free to get started. Docker's product page leads with a "Get started free" call to action. Docker Business, the paid tier for team security and compliance, costs $24 per user per month; Docker AI Governance is priced through a sales form.

Do I need Docker Desktop to use Docker Sandboxes?

No. Docker's FAQ states sandboxes do not require Docker Desktop. Installation is a standalone CLI: brew install docker/tap/sbx on macOS, winget install Docker.sbx on Windows, or apt-get install docker-sbx on Ubuntu.

How do Docker Sandboxes compare with E2B and Daytona?

Docker Sandboxes runs free local microVMs on your machine. E2B runs cloud sandboxes on usage-based pricing: Hobby at $0 per month with $100 in credits, Pro at $150 per month. Daytona charges per hour with $200 in free compute, including GPU instances from $0.99 per hour.