Dozable blog

Dozer Sandbox 0.35 - doz try, and keys from any provider

2026-10-18T00:00:00Z

It's on the stable channel now, as 0.35.1. It includes everything from 0.34, which went straight from canary to this release.

doz try#

One command puts you in a throwaway sandbox with any of 42 coding agents and AI tools: its own empty folder, a network that reaches only its model, and your key kept on your Mac. doz try has its own post.

doz try opencode

Keys from any provider, and a choice of model#

Use a key from OpenAI, Anthropic, Google or xAI, or one key for many models through OpenRouter or Vercel AI Gateway. Dozer can also read a key from 1Password, AWS Secrets Manager or a command each time it's needed. A try's setup and New Sandbox let you pick the model from the gateway's own list, and doz models lists them.

The Accounts & keys page groups keys by provider, with the tools each one serves. A key that stopped working comes first, with what the provider said and how to fix it.

Tools#

OpenCode 2 is the opencode Dozer recommends; OpenCode 1 stays as opencode@1. Its free OpenCode Zen models need no key.
Grok Build, Factory Droid and Docker Agent report working and done in the dashboard.

Also new#

doz search aider says whether Dozer runs a tool, and how, or why not yet.
A sandbox's web page opens in your Mac's browser: the sandbox's Open button, or doz open.

Fixed#

Icons in the dashboard no longer blink every few seconds.
A key its provider refuses shows as refused, instead of a green "ok".

Upgrade#

doz upgrade

doz try - any of 42 AI tools in one command

2026-10-18T00:00:00Z

Pick a tool and Dozer makes the sandbox, starts it and puts you in it. Nothing is installed on your Mac, and nothing of yours is shared with it. The first try of a tool prepares its image, which takes a few minutes, once.

doz try                 # the list: coding agents, then AI tools, and which are ready
doz try opencode        # OpenCode, in a throwaway sandbox
doz try fabric --rm     # delete it when you leave

What a try gets#

A new empty folder under your projects folder, shared at /workspace, so what the tool makes is on your Mac. Never one of yours.
A network that reaches its model and nothing else. Every other host is refused and logged.
Your key, kept on your Mac. The sandbox only ever holds a placeholder. No key yet? The try asks for one, once.

--dry-run shows all of it before anything is made: the version, every host it can reach, the folder and the key.

doz try grok --dry-run

When you leave, it asks whether to sleep it, delete it or keep it running. Deleting it removes its folder too, and asks first if the tool left files there.

42 tools#

Claude Code, Codex and pi, plus Gemini CLI, Grok Build, OpenCode, GitHub Copilot CLI, Goose, Qwen Code, Cline, OpenHands, Crush and more coding agents. And AI tools such as Fabric, llm, AIChat, ShellGPT, promptfoo and Google Stitch. The manual lists them all, with what each needs.

Choose the model#

Use a key from OpenAI, Anthropic, Google or xAI, or one key for many models through OpenRouter or Vercel AI Gateway. OpenCode can also use OpenCode Zen's free models, with no key at all.

doz try opencode --gateway openrouter --model anthropic/claude-sonnet-4.6
doz try opencode --provider opencode-zen

In the dashboard#

Try → Quick try shows the same tools as cards. Search, or filter by kind and category, and Try now opens its setup: where it gets its model, your key and the folder.

Each try opens on its Guide: what the tool is, the facts of this try, and its first commands. Run types a step into the tool's session; Copy puts it on your clipboard.

doz try is in Dozer Sandbox 0.35. Upgrade with doz upgrade, then doz try.

OpenCode 2 is in Dozer Sandbox

2026-10-17T00:00:00Z

OpenCode 2 is out, and you can try it in a sandbox with one command. OpenCode 1 stays available, so a project that depends on it keeps working.

doz try opencode        # OpenCode 2, in a throwaway sandbox
doz try opencode@1      # OpenCode 1

It needs Dozer Sandbox 0.35, now on the stable channel.

In the dashboard, OpenCode's card in Quick try has a Version choice; it starts on 2. Sandboxes you already have stay on the version they were made with.

What's new in OpenCode 2#

Anomaly, the company behind OpenCode, announced 2.0 on 12 October as "a major overhaul" (their post), about three months after its public beta. Its changelog has moved to the 2.x series, so 2 is now OpenCode's main release (The New Stack has a good account of the release).

One service, several interfaces. OpenCode now starts a single background service. The terminal interface, the web app and the desktop app all talk to it, so you can start a session in the terminal and continue it in the browser. Your own scripts can find and use the same service.
A new API that lets the agent manage itself. The agent can create a git worktree for a new task, move a session into it, or find and interrupt a session that is stuck. Scripts reach the API through @opencode/client; @opencode/sdk embeds OpenCode in your own application (SDK docs).
Code Mode. Instead of one tool call per round trip to the model, the agent writes a short JavaScript program that calls several MCP tools or APIs and works with the results, for example reading bug reports from one service and filing issues in another. It is the default for every MCP server.
Work in the background. Subagents and shell commands can run while you keep working with the main agent: a code review by a subagent on another model, or a long test run that reports back when it finishes. A shortcut (ctrl+b) moves a running task to the background.
Plugins. More than 80 of its own features are now plugins, terminal-interface plugins are stable, and plugins hot-reload, so OpenCode can write a plugin and use it in the same session (how they rebuilt it).

What breaks: plugins written for OpenCode 1 don't load in 2, and integrations built on the old server API need updating.

In Dozer#

OpenCode 2 gets what every tool in Dozer gets:

Your key stays on your Mac. Use a key from xAI, OpenAI, Anthropic, OpenRouter or Vercel AI Gateway; the sandbox only ever holds a placeholder. OpenCode Zen's free models take no key at all.
A locked network that reaches its model and nothing else, and a folder of its own. OpenCode 2's background service runs inside the sandbox and needs nothing more.
A guide beside the try with the first commands to run.
Its web app on your Mac: run opencode pair in the sandbox and open the link it prints, or use the sandbox's Open button; it shows in your browser, on this Mac only.

Outside Dozer, OpenCode 1 users move with opencode upgrade --major or /update; since 1.19 it no longer updates itself (1.19.0 release notes). In Dozer you choose the version by name instead, and the sandbox's version comes from Dozer's checked recipe.

Ways to try it#

OpenCode Zen's free models need no account and no key. The picker and doz models --provider opencode-zen show each model's data terms; some may use what you send to improve their models.

doz try opencode --provider opencode-zen          # Big Pickle, free, no key

Or your own key from xAI, OpenAI or Anthropic, or one key for many models through OpenRouter or Vercel AI Gateway:

doz try opencode                                  # OpenCode 2 on a key you've added
doz try opencode --provider anthropic             # on your Anthropic key
doz try opencode --provider openai                # on your OpenAI key
doz try opencode --gateway openrouter             # through OpenRouter
doz try opencode --gateway openrouter --model anthropic/claude-sonnet-4.6
doz try opencode --gateway vercel-ai-gateway      # through Vercel AI Gateway
doz try opencode@2.0.29                           # exactly this release
doz try opencode --rm                             # delete it when you leave

No key for that provider yet? The try asks for one, once, or add it first with doz account add openrouter --provider openrouter. Add --dry-run to any of these to see what it would make and what it can reach, without making anything. In the dashboard, Try → Quick try → OpenCode has the same choices: the Version, and where OpenCode gets its model, with a list of a gateway's models to search.

Why OpenCode 1 stays#

A major version changes what your setup depends on: here, plugins and the API. Dozer pins each tool to a version it has checked. opencode now means OpenCode 2, the version Dozer recommends; OpenCode 1 stays one word away, as opencode@1, and keeps getting 1.x updates. An existing sandbox never moves to a new major version by itself.

New to Dozer Sandbox? Getting started takes a few minutes.

Grok Build now tells Dozer what it's doing

2026-10-17T00:00:00Z

Grok Build 1.0.52 reports its status, and Dozer Sandbox 0.35 runs that version. Start a Grok Build sandbox and leave it to work; the dashboard shows when it's busy, when it needs you, and when it's finished.

doz try grok

What you see#

While Grok Build works, its tab says working, and a dot beside the sandbox in the sidebar says something is going on.

When it stops to ask you something, such as permission to run a command, a notice at the top of the dashboard says so, with an Open button that takes you straight to the session, whichever page you're on.

When it finishes, the tab says done. All sessions shows the same on every tile, so with several agents running you can see which one is still busy and which is waiting for you.

Where it comes from#

xAI's changelog for 1.0.52 lists "support for OSC 7501". That's a terminal escape sequence for program status, proposed by Mitchell Hashimoto: a program asks the terminal whether it understands the sequence, and if it does, reports its state each time it changes.

Inside every sandbox, Dozer answers that question for each session and keeps the latest report, whether or not the dashboard is open. Nothing is read from the screen or guessed: Dozer shows exactly what Grok Build reports.

Agents that report their status#

Claude Code, pi and now Grok Build report their status. So do Factory Droid and Docker Agent, which also arrive in 0.35. When another tool adds support, Dozer turns it on in that tool's recipe once we've seen it work.

Upgrade to 0.35, then doz try grok.

Introducing Dozer Sandbox

2026-10-11T00:00:00Z

Coding agents are at their best when you stop approving every command. They're at their most worrying then too: the agent can read your files, run anything, and reach any site, all with your keys in its environment.

Dozer Sandbox gives each agent its own small Linux computer on your Mac, so you can let it work.

What a sandbox is#

Each sandbox is a small Linux virtual machine with one project folder from your Mac inside it, at /workspace. Run Claude Code, Codex or pi in it with the permission prompts off: the VM is the boundary.

Fast. A sandbox starts in about a third of a second and pauses in a millisecond.
Resumable. Put it to sleep and it gives its memory back. Wake it and every program is still running where you left it, mid-sentence.
Secure. The network goes through a policy you can read, and your API keys never go inside: the agent sees a placeholder, and the real key is added on your Mac on the way out.

Try it#

You need a Mac with Apple silicon and macOS 26 or later, and Homebrew.

brew trust --tap dozer-sandbox/tap
brew install dozer-sandbox/tap/doz
doz ui

The dashboard opens on a short setup wizard: it checks your Mac, asks how your agent should sign in, and prepares an image. Then New sandbox and you're in. Prefer the terminal? doz onboard, then doz up in a project folder.

Free and open source#

Dozer Sandbox is free, and its source is on GitHub.

This blog is where news, early access and tips will appear. They're also sent by email to anyone who signs up for them. Thanks for reading, and for trying it.

Dozer Sandbox 0.33 - see what your agent is doing

2026-10-11T00:00:00Z

It's on the stable channel now, as 0.33.1. Here's each change.

See what the agent is doing#

Claude Code and pi now tell Dozer what they're doing: working, done, waiting for your answer, or stopped by an error. Dozer shows it in the dashboard wherever the session appears, so you can leave an agent to work and glance back when it matters.

While the agent works, its tab shows working, and a dot beside the sandbox in the sidebar says something is going on.

When the agent stops to ask you something, such as permission to run a command, a notice appears at the top of the dashboard saying what it wants, with an Open button that takes you straight to the session. You see it on whichever page you're on.

When it finishes, the tab says done and the notice tells you so. The same status shows on every tile in All sessions, so with several sandboxes running you can see at a glance which agent is still busy and which is waiting for you.

The status shows while the dashboard is open (doz ui). Claude Code needs version 2.1.295 or later, and pi 1.1.0 or later.

How it works#

It uses a terminal escape sequence: bytes mixed into a program's output that the terminal acts on instead of showing, the same way programs set a window's title or colours. This one is OSC 7501, for program status (Mitchell Hashimoto's write-up).

The handshake: a program asks the terminal whether it understands the sequence (ESC ] 7501 ; ? ST). If the terminal answers, the program reports its state (working, done, blocked or error, with a short message) each time it changes.
Inside the sandbox: every session runs under a small holder process that reads everything the program writes, whether or not anyone is watching. It answers that question, keeps the latest report, and passes the output on unchanged.
On your Mac: the Dozer host watches each running session's status, and the dashboard draws it.

Nothing is read from the screen or guessed: Dozer shows exactly what the agent itself reports.

`doz upgrade`#

One command moves you to the newest version on your channel. It asks first; -y skips the question.

doz upgrade

The dashboard in the background#

doz ui --detach runs the dashboard without holding a terminal, and doz ui stop stops it from any terminal.

Staying in touch#

Setup now ends with an optional sign-up for release news, early access, and tips and tricks, by email. Skip it, or sign up later with doz signup.

Fixed in 0.33.1: the agent can write your workspace again#

Since 0.30, an agent in a sandbox could read a shared project folder but not change it, and git refused the folder ("dubious ownership"). 0.33.1 fixes it. After upgrading, a sandbox that was running, asleep or hibernated needs a doz shutdown NAME and then doz start NAME; one that was shut down picks up the fix at its next start.

Upgrade#

doz upgrade

New to Dozer Sandbox? Getting started takes a few minutes.

Dozable blog

Dozer Sandbox 0.35 - doz try, and keys from any provider

doz try#

Keys from any provider, and a choice of model#

Tools#

Also new#

Fixed#

Upgrade#

doz try - any of 42 AI tools in one command

What a try gets#

42 tools#

Choose the model#

In the dashboard#

OpenCode 2 is in Dozer Sandbox

What's new in OpenCode 2#

In Dozer#

Ways to try it#

Why OpenCode 1 stays#

Grok Build now tells Dozer what it's doing

What you see#

Where it comes from#

Agents that report their status#

Introducing Dozer Sandbox

What a sandbox is#

Try it#

Free and open source#

Dozer Sandbox 0.33 - see what your agent is doing

See what the agent is doing#

How it works#

doz upgrade#

The dashboard in the background#

Staying in touch#

Fixed in 0.33.1: the agent can write your workspace again#

Upgrade#

`doz upgrade`#