Sandbox Console

Run Claude Code, Codex or Gemini inside a real microVM — from a VS Code terminal that feels native.

• Isolated — the agent runs in a Docker Sandbox (sbx) microVM, not on your host.
• Persistent — stop and resume keeps authentication, installed packages, and history.
• Zero container management — one Explorer tree, no YAML by hand, no docker commands.

Install

Search Sandbox Console in the VS Code Marketplace, or:

• Quick Open (Ctrl+P): ext install ainova-systems.sandbox-console
• Terminal: code --install-extension ainova-systems.sandbox-console

Then open a repository and run Sandbox: New Sandbox (or click + New Sandbox in the status bar) → pick an agent → Connect. A terminal opens with the agent running inside the sandbox, in your workspace.

Features

• Sandboxes view — an Explorer tree of this repo's sandboxes with live status. Per-node actions are state-gated (running → Stop → Edit / Rebuild / Delete instance → Remove from config) and can be organised into groups.
• Status bar & sandbox switcher — the active sandbox and its state are always visible; one click connects. Switch Sandbox picks another; the committed recipe (.sandbox/config.yaml, see Configuration) can mark one default.
• New / Edit form — a webview panel for agent, title, group, credentials, ports, and environment. Saving writes the recipe and applies to the instance.
• Custom environments — per sandbox: the agent's default image, a custom Dockerfile (a starter is generated, then docker build → sbx template load), or a registry image used as-is.
• Credential provisioning — tick the services a sandbox needs; values are piped over stdin to sbx secret set — never written to the repo, an image, or a shell history. Per-project values are cached (encrypted for your OS user) so a repo-scoped token is entered once, not per sandbox. For github the gh CLI is authenticated inside too.
• Generated project CLI — .sandbox/scripts/sbx.sh exposes the same sandbox model (status, connect, exec, headless tasks, ephemeral runners, secrets) to shells, scripts, and AI skills, so automation never has to re-encode the naming and lifecycle rules.
• Multi-agent, multi-sandbox — Claude Code, Codex, Copilot, Cursor, Gemini, OpenCode, Droid, Kiro, plain shell, and whatever else your sbx reports (the list is discovered live). Parallel work on one repo uses git worktrees or mount: clone.

Commands

Command Palette, category Sandbox:

Long operations (image builds, creates) stream their output into both the progress notification and the log, and can be cancelled from the notification. Only one operation runs per sandbox at a time — a second click reports what is already running instead of starting a duplicate.

Explorer per-node actions: Connect, Stop, Shell, Rebuild, Edit, Delete instance, Remove from config.

Requirements

• VS Code 1.90+
• Docker Sandboxes (sbx), installed and signed in with sbx login. Per Docker's documentation it is available for macOS 14+ (Apple silicon), Windows 11 (x86_64, Windows Hypervisor Platform enabled), and Ubuntu 24.04+ (KVM); the docs carry no beta or limited-access label, and the CLI is free to use including commercially. Docker Desktop is not required by sbx itself.
• Docker Engine / Docker Desktop on the host — only for the custom Dockerfile mode, which runs docker build locally.
• Workspace Trust — the extension runs sbx commands derived from the workspace's committed .sandbox/config.yaml, so it stays inert in untrusted and virtual workspaces.
• An account for whichever agent you run (e.g. Claude Code's /login inside the sandbox).

Verified against sbx v0.31.3 (2026-07-28), Windows 11.

Platform support: developed and verified on Windows (sbx is resolved from %LOCALAPPDATA%\DockerSandboxes\bin). macOS and Linux are untested — sbx is taken from PATH and POSIX workspace paths pass through unchanged, so the core flows are expected to work, but the per-project secret cache is Windows-only (DPAPI) and elsewhere degrades to prompting for each value. Workspaces on UNC / \\wsl$ paths are not supported. Reports welcome.

Configuration

Sandboxes live in a committed, compose-like recipe written by the form:

version: 1
name: my-repo
sandboxes:
  claude:
    agent: claude
    title: Backend
    group: Services
    secrets: [github]
    ports: [5173]
    default: true
  dotnet:
    agent: shell
    image: myrepo-dotnet:latest
    dockerfile: dotnet.Dockerfile   # built and tagged as `image`

What the extension writes into your repository — only once you create your first sandbox, never on merely opening a project:

How it works & security

The extension is a thin orchestration layer over the sbx CLI — it does not manage containers itself. Your workspace is mounted into the microVM automatically, and credentials are proxied host-side, so a per-sandbox secret is never exposed to another sandbox's environment (globally scoped credentials are deliberately shared). Secret values travel only over stdin pipes.

The generated .sandbox/scripts/sbx.sh dispatches headless agent tasks with claude --dangerously-skip-permissions — that auto-approval is deliberate and confined to the isolated sandbox, where the blast radius is the microVM and the mounted workspace, and it is never used for host-side execution. The extension generates that script but never executes it.

Security policy and vulnerability reporting: SECURITY.md.

Links

• Features and Architecture
• Iteration specs
• Contributing — build, debug, and PR conventions
• Issues & support
• Upstream: Docker Sandboxes documentation

AI-First engineering, as a service

Sandbox Console was designed and implemented by AI coding agents, end to end — working proof of the method, not a claim about it. Ainova Systems helps teams adopt the same way of working: AI automation, an AI-First SDLC, and fully autonomous engineering pipelines — with humans owning the decisions.

Talk to us →

Licensed MIT. © MB Ainova Systems — https://www.ainovasystems.com

Docker is a trademark of Docker, Inc. Claude is a trademark of Anthropic, PBC. This project is independent and is not affiliated with, endorsed by, or sponsored by Docker or Anthropic.