Sandbox ADE

The Sandbox ADE (Agentic Development Environment) is a native desktop app for macOS and Linux that runs coding agents inside Tenki Sandbox microVMs. Start a chat, hand the task to an agent, and watch it work inside a disposable Linux VM with real root while your machine stays untouched.

The ADE exists because agents are most useful exactly when they're most dangerous: when they run real commands. Instead of giving an agent root on your laptop, the ADE gives every task its own isolated Firecracker microVM, full visibility into everything the agent does there, and explicit permission prompts before risky actions.

If you'd rather drive sandboxes from code or a terminal, use the SDK or CLI. All surfaces talk to the same microVMs.

Download and install

Platform	Package	Download
macOS (Apple silicon)	.dmg	Download for macOS
Linux x64 (Debian/Ubuntu)	.deb	Download .deb
Linux x64 (AppImage)	.AppImage	Download AppImage

There is nothing else to set up locally: no Docker, no local virtualization. Sandboxes run in Tenki Cloud, so the app only needs network access plus git on your machine. The GitHub CLI (gh) is optional but recommended for pull-request flows.

First launch

Setup takes about two minutes:

1. Sign in to Tenki Cloud. Choose Continue with Tenki Account to authenticate in your browser, or paste a Tenki API key if you prefer. Then pick the workspace and project your sandboxes should run (and bill) in. No account yet? Sign-up is part of the same flow.
2. Setup check. The app verifies git is installed (required) and looks for the GitHub CLI (optional).
3. Connect an agent. Sign in to at least one agent provider (see the table below). You can add or change providers later in Settings → Providers.

Supported agents

The ADE ships three agent harnesses. Each one signs in with the subscription you already have, or a provider API key:

Agent	Subscription sign-in	API key alternative
Claude Code	Claude account (claude CLI login or a service token)	Anthropic API key
Codex	ChatGPT account (codex login)	OpenAI API key
Pi	Pi account	Anthropic and OpenAI API keys

Per chat, the composer toolbar lets you pick:

• Model: current models from each provider; picking a model selects its agent.
• Effort: reasoning level, where the model supports it.
• Mode: the agent's permission mode, from read-only and plan modes through "ask before risky actions" (the default) up to full auto-accept.

Start a task

A new chat asks three things:

1. Project and branch. Pick a GitHub repository (add one with Mod+Shift+O) and the branch to start from, or no repo at all for a blank sandbox. The ADE clones the repo into the sandbox using your GitHub credentials.
2. Where it runs. Cloud (the default) creates a fresh sandbox microVM for the chat. You can also run a chat locally, in a git worktree or directly in your local clone, when you explicitly want the agent on your machine instead of in a VM.
3. Environment. By default the sandbox boots from the base image. Pick a blueprint (a pre-built environment image or a snapshot) instead, and customize the machine size (CPU, memory, disk). Repositories can save a default blueprint and size so new chats are one keystroke.

Then type the task and send. The agent works inside the VM with root access: it can install packages, run servers, and execute whatever the task needs without touching your host.

While the agent works

Everything the agent does streams into the chat as it happens: messages, thinking, file edits, shell commands, and tool calls.

• Permission prompts. In the default mode, the agent stops and asks before risky tool calls. You approve once, always-allow for the session, or deny. Switch the mode per chat if you want plan-only, auto-accept edits, or full autonomy.
• Review diffs (Mod+E). A change review panel shows uncommitted and committed changes per file, with side-by-side or unified diffs and line comments you can send back to the agent.
• Terminal (Mod+J). A real shell inside the sandbox: inspect state, run commands, or fix things yourself while the agent works.
• Files panel (Mod+G) to browse and search the workspace.
• Embedded browser. Open the app the agent is building in a built-in browser pane. An annotate mode lets you click an element on the page and drop it, screenshot and styles included, straight into your next message.
• Preview URLs. Agents can expose a port from the sandbox and hand you a public preview URL for the dev server they just started.
• Notifications. Optional native notifications when an agent needs approval, errors, or finishes, with per-event toggles and a long-running-task threshold in Settings.

Because the agent is in a disposable VM, a bad command (an errant rm -rf, a prompt-injected curl-pipe-bash) has zero blast radius. Throw the session away and start clean.

Ship the result

Code leaves the sandbox through git, and you hold the controls. A context-aware button in each chat offers the next sensible step (Commit, Commit & push, or Commit, push & create PR) with AI-drafted commit messages and PR descriptions you can edit before they go out. The PR opens on GitHub in your browser.

Session lifecycle and billing

ADE cloud chats are regular sandbox sessions. They show up in your workspace alongside sessions created through the CLI or SDKs, follow the same lifecycle, and bill the same way; see Sandbox pricing.

• Reopen to resume. The sandbox keeps running when you switch away or quit the app. Reopen a recent chat and the ADE reconnects to the same VM in seconds, with the full transcript restored.
• Sessions auto-extend while in use. Sessions have a max duration (default 2 hours, configurable from 30 minutes to 24 hours in Settings). The ADE extends it automatically while the agent is working or you're active, and warns you before an idle session expires.
• Archive to clean up. Archiving a chat (Mod+Backspace) terminates its sandbox but keeps the transcript, so you can still read (and unarchive) the conversation later.
• Snapshot to keep VM state. Take a snapshot of the sandbox (Mod+Shift+S) and start future chats from it. This is handy for preserving a fully warmed-up environment before archiving.

Keyboard shortcuts

Press Mod+/ in the app for the full list (Mod is Cmd on macOS, Ctrl on Linux). Highlights:

Shortcut	Action
Mod+K	Command palette
Mod+E	Open review diff
Mod+J	Toggle terminal
Mod+G	Open files panel
Mod+D	Split pane side-by-side
Mod+Shift+I	Change model
Mod+Shift+M	Change permission mode
Mod+Shift+S	Snapshot the sandbox
Mod+Shift+G	Open pull request
Mod+Backspace	Archive chat

Settings

Settings (Mod+,) covers, among other things:

• General: theme (system, light, dark, black), terminal and code fonts, the model used for drafting commit messages and PR text, smart chat titles, default sandbox size and session duration, and the update channel (stable or beta).
• Providers: agent sign-ins and API keys for Claude Code, Codex, and Pi; enable or disable individual providers.
• Repositories: manage your projects and their blueprints, and set the default blueprint for new chats.
• Notifications: the global toggle plus per-event settings.

What's next

• New to Tenki Sandbox? Read the Concepts page for how sessions, volumes, and snapshots fit together.
• Prefer code? The Quickstart covers the CLI and the TypeScript, Go, and Python SDKs.
• Hit a problem? Check Troubleshooting or email hello@tenki.cloud.