Docker rebranded for AI agents, but Sandboxes shipped back in September
On November 19, Docker changed its brand and slogan to "Run on Trust," but its agent products had been rolling out all year. The most useful of them is Sandboxes: Claude Code or Codex runs in a separate microVM with its own kernel, so a deleted folder won't touch your laptop. The sbx CLI and local sandboxes are free; only the enterprise policy consoles are paid. Developers should try sbx rather than look at the logo.
AI-processed from Habr AI; edited by Hamidun News
On November 19, Docker changed its brand: the whale was redrawn, the blue color was swapped, and the homepage gained the slogan "Trust Docker for the agents you don't." It looks like the launch of a whole lineup of new products for AI agents. But if you check the New labels on the site against release dates, the picture is different: Docker released almost all of these over the course of a year and in November simply moved them to the front of the store. For practitioners, what matters more than the logo is what from this set you can install today.
What is actually on the menu
The New label appears on three products. Sandboxes was discussed on Hacker News back in the summer, and cloud sandboxes launched on September 24. AI Governance was mentioned in July. The open-source MCP Gateway was already in Docker Desktop, and MCP Enterprise Gateway is its enterprise version. Hardened Images have been free since December 17, 2025, while Hub, Desktop and Offload are old acquaintances.
The showcase shows who is being sold what. Developers are offered a sandbox to install. Companies are offered a conversation about who will manage agents' access and how. The rebranding text uses the word "trust" more than twenty times, while containers got a single paragraph promising further development.
How Sandboxes work
The problem is a familiar one: you ask Claude Code to update dependencies, and it decides to delete the folder and rebuild everything from scratch. If the agent runs directly on your laptop, its commands can affect everything it has access to. By default, Claude Code also starts with the --dangerously-skip-permissions flag and does not ask for permission for each deletion.
A sandbox gives the agent a separate environment. Each one runs in its own microVM with its own kernel and Docker Engine. A regular container shares the kernel with the host, whereas here a hypervisor sits between the agent and the machine: KVM on Linux or Windows Hypervisor Platform on Windows. You only need the sbx CLI, and launching looks like this: sbx login, then sbx run claude. Instead of claude, you can specify codex, copilot, cursor or opencode.
By default, the current folder is mounted into the sandbox with write access, meaning the agent can change it directly on the host. The --clone flag runs it in a separate clone of the repository: if the experiment goes wrong, you won't have to clean up your working copy. Additional folders can be attached as read-only, for example ~/shared-libs:ro.
Secrets are handled much as they are in frontend development: the browser sends a request to the backend, and the backend substitutes the key. The agent in the sandbox sees a proxy-managed placeholder, while the real token stays on the host, and the proxy inserts it into the request header. The agent cannot read the token, but it can send requests into which the proxy will insert it. Where those requests go is decided by the network policy: the required domain is added with a command such as sbx policy allow network registry.npmjs.org.
There are limitations. macOS requires Apple silicon and Sonoma 14+, and on Linux Docker specifies Ubuntu 24.04+. Sign-in via sbx login is required, and the Hacker News discussion argued specifically about tying it to an account and about the placeholder in place of the key. According to the FAQ, the CLI and local sandboxes are free, including for commercial use.
The paid part and why Docker is doing this
AI Governance is a console where a platform team hands out policies to agents across the whole company: which domains are open, which folders can be mounted, which MCP tools can be called. Audit logs go to Splunk, Datadog or Grafana. If an organization has enabled centralized management, the developer's local permissions stop applying. MCP Enterprise Gateway authenticates the user through corporate SSO, decides which MCP servers they are entitled to, and injects credentials itself. Both products are priced "on request" or behind a "Talk to sales" button.
For an individual, the open-source MCP Gateway under the MIT license is enough. Hardened Images in their basic variant are also free: more than a thousand stripped-down images based on Debian and Alpine under Apache 2.0. Paid variants add FIPS, guarantees on vulnerability fix timelines, and extended support. All of this is needed when hundreds of employees run agents and someone has to report to the information security department.
The company's motive is clear. Since 2025 it has been led by Don Johnson, who previously built cloud infrastructure at Oracle, and after his arrival analysts discussed a possible sale of the business, which Docker did not comment on. Podman, containerd, OrbStack and any cloud provider can run containers today, so growing on a single docker run is hard. At the same time, security teams have a new question: how to give an agent a terminal without losing production. Docker has sold code isolation for ten years, and this is its territory.
Market interest is visible in the numbers. The Hacker News thread about the rebrand gathered eight points and a comment calling it "vacuous," while the summer discussion of Sandboxes reached nearly 700 points and 400 comments. The engine, the images and docker compose up haven't gone anywhere, but the first screen of docker.com is now entirely about agents.
The practical takeaway is simple. If you run Claude Code or Codex with confirmations disabled, try sbx with the --clone flag: it's free and takes almost no time. Enterprise consoles are worth discussing when you have many agents and need auditing. The open question is whether Docker will find a new business or simply adapted to the hype, but the sandboxes themselves can already be tested on your own projects.
- Nov 21, 2026Google Releases Gemini 4 Neon: 65.2% on Terminal-Bench 4.1 at $3 per Million Tokens
- Nov 21, 2026Skein Agent Erases Finished Stages From Context: 97% Success Rate, But More Expensive Than OpenCode
- Nov 21, 2026CoreWeave Owes $47 Billion on Nvidia Chips and Lost $709 Million in One Quarter
- Nov 21, 2026ClawDOS Writes and Runs Code on an IBM PC 5160: Primes Under 100 in 64 Seconds
Key Questions Answered
What is actually on the menu?
The New label appears on three products. Sandboxes was discussed on Hacker News back in the summer, and cloud sandboxes launched on September 24. AI Governance was mentioned in July. The open-source MCP Gateway was already in Docker Desktop, and MCP Enterprise Gateway is its enterprise version. Hardened Images have been free since December 17, 2025, while Hub, Desktop and Offload are old acquaintances. The showcase shows who is being sold what. Developers are offered a sandbox to install.
How Sandboxes work?
The problem is a familiar one: you ask Claude Code to update dependencies, and it decides to delete the folder and rebuild everything from scratch. If the agent runs directly on your laptop, its commands can affect everything it has access to. By default, Claude Code also starts with the --dangerously-skip-permissions flag and does not ask for permission for each deletion. A sandbox gives the agent a separate environment. Each one runs in its own microVM with its own kernel and Docker Engine.
What are the key details on The paid part and why Docker is doing this?
AI Governance is a console where a platform team hands out policies to agents across the whole company: which domains are open, which folders can be mounted, which MCP tools can be called. Audit logs go to Splunk, Datadog or Grafana. If an organization has enabled centralized management, the developer's local permissions stop applying. MCP Enterprise Gateway authenticates the user through corporate SSO, decides which MCP servers they are entitled to, and injects credentials itself.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.