A sandbox platform from Docker, Inc. for running AI coding agents safely inside isolated microVMs. The agent works in a disposable environment fully cut off from your host, where it can freely perform file operations, install packages, and even launch containers with root privileges — so it never has to stop and wait for your approval. Released at the end of March 2026 as the sbx CLI tool, it supports the major coding agents including Claude Code, Codex, Gemini CLI, and GitHub Copilot CLI.
Key Features
- Strong isolation via microVMs: Each sandbox gets its own dedicated Docker daemon, filesystem, and network, and does not share the host kernel. This gives you a firmer isolation boundary than a regular container.
- YOLO mode: Flags like
--dangerously-skip-permissionslet the agent run autonomously without approval prompts, so work proceeds without a confirmation step at every turn. - Sandbox Kits: Declare environment configuration and permissions in YAML, then distribute and reuse them across your team as reproducible work environments.
- Docker inside the sandbox: The agent can build and launch independent containers inside the sandbox without mounting the host’s Docker socket.
- MCP gateway: MCP servers registered on the host are bridged centrally to the agents inside the sandbox (the
sbx mcpcommand family). This saves you from configuring each agent individually. - Protection through a network proxy: Traffic is relayed on the host side, blocking access to the host’s localhost while automatically injecting authentication headers.
Pricing
| Plan | Monthly Price | Key Features |
|---|---|---|
| Free | $0 | sbx CLI and sandbox execution available at no cost |
Docker Sandboxes itself is free and does not require Docker Desktop. If you need centralized management such as organization-wide network policies or MCP governance, that lives in a separate product, Docker AI Governance (contact sales). It sits apart from Docker’s general subscriptions (Personal/Pro/Team/Business), and no paid plan dedicated to Sandboxes appears on the official site.
Pricing reflects information as of August 2026. Check the official site for the latest details.
Pros and Cons
✅ Pros
- microVM-level isolation means less exposure for your host than a regular container, making it well suited to running untrusted AI-generated code
- The combination of YOLO mode and Sandbox Kits gives you both approval-free autonomous execution and reproducible team environments
- Broad support for the major agents: Claude Code, Codex, Gemini CLI, Copilot CLI, and more
- Runs on Mac, Windows, and Linux, with easy installation through package managers
⚠️ Cons
- There’s no standalone pricing page for Sandboxes, and pricing for the organization-facing product (Docker AI Governance) requires contacting sales
- Accounts of the official release timeline vary — it rolled out in preview in late 2025 and reached GA in stages through 2026 — so you’ll want to track ongoing updates to gauge feature maturity
- Sandboxes assume local execution, a different design philosophy from platforms built primarily for large-scale parallel execution in the cloud
Comparison with Similar Services
| Criteria | Docker Sandboxes | E2B | Daytona |
|---|---|---|---|
| Isolation model | microVM (dedicated kernel, independent Docker daemon) | Firecracker microVM (dedicated kernel per session) | Docker container (shares host kernel) |
| Primary use case | Autonomous AI agent execution on your local machine | Short-lived code execution (function-call style usage) | Agent execution in a persistent development environment |
| Pricing | Free (organization features are a separate product, contact sales) | Usage-based | Usage-based |
| Supported agents | Claude Code / Codex / Gemini CLI / Copilot CLI, etc. | Embeddable in any agent via SDK | Embeddable in any agent via SDK |
Who It’s For
- Developers who want to hand autonomous work to AI coding agents like Claude Code or Codex without approval prompts
- Engineers who want to try package installs and container operations without polluting their host environment
- Teams who want to define their agents’ work environment in YAML and share it in reproducible form
- Users who want to connect multiple MCP servers to their agents from one place
Summary
Docker Sandboxes is a platform that came into its own in 2026, supporting autonomous AI coding agent execution on two fronts: isolation and speed. It combines the strong separation of microVMs with the autonomy of YOLO mode, and being free to use lowers the barrier to adoption. Once you reach the point of needing organization-wide governance, it’s worth keeping the separate Docker AI Governance product on your radar.