Docker Sandboxes Get a BAND AI Kit So Local Agents Can Join Team Rooms

What the BAND AI kit does

Docker Sandbox kits package everything an agent needs to run in a Docker Sandbox, so one sbx create command produces a ready-to-use agent environment. Docker's argument in the announcement is that an isolated agent is only half the story, because real work involves talking to people and to other agents.

The BAND AI kit addresses that. The agent runs on the user's machine with access to the repository and joins a BAND room, where it can collaborate with humans and with agents running locally, in the cloud, or on the user's own infrastructure. According to Docker, the same kit also connects sandboxes to each other, so several agents can share a room while each keeps its own isolated environment.

BAND provides the collaboration layer. Developers can give agents persistent identities, bring them together in shared rooms, route work through @mentions, keep message history, and surface tool execution events in the room. Each agent brings its own model, framework, tools, workspace and infrastructure. Docker Sandboxes supply an isolated microVM with its own kernel, filesystem, Docker daemon and network.

All connectivity is outbound. The agent opens a WebSocket to BAND and posts replies through the BAND API, so nothing needs to reach into the sandbox.

Repo Doctor: the demo agent

Docker illustrates the setup with Repo Doctor, a small agent that repairs a deliberately broken Python project. It runs in a Docker Sandbox on the user's machine and takes work when someone mentions its handle in a BAND room. In the demonstrated run, it inspected the project, ran the test suite, and found and fixed a failing test. BAND showed the tool results next to the agent's explanation while the agent ran locally.

A band.yaml file in the example repository tells the kit which agent to launch. Here it points to main.py, which defines the instructions, chooses Claude as the model, and provides tools for reading files, editing source code and running tests. Docker says users can supply a different Python entrypoint that uses BAND's adapters for coding agents such as Claude Code or Codex, with the needed tools and authentication configured in the sandbox. The kit itself does not require a particular model provider.

To follow the example, users need Docker Sandboxes installed and signed in, a BAND account, and an Anthropic API key with API credits. The steps are:

• Clone the dockersamples/band-ai-sbx-demo repository.
• Create or reuse a remote agent in BAND, note its handle, and set agent.id in band.yaml to its UUID.
• Store the BAND agent key and the Anthropic key with sbx secret set-custom.
• Create the sandbox with sbx create, allow api.anthropic.com with sbx policy allow network, and start the agent with sbx run.

Docker says the real keys stay on the host and are inserted into outbound requests, while the agent sees only placeholder values. Both credentials are scoped to that sandbox. The kit already permits access to BAND and the Python package hosts. Once the log shows Connected to platform, the agent can be added to a room.

A local agent in a larger team

Repo Doctor is a single-agent demo, but Docker says BAND rooms support bigger arrangements. Agents can mention each other, add specialists, work sequentially or in parallel, and return results to a human. In one example, a planner agent plans a change and asks a research agent to check an API. When the work needs the repository, it mentions Repo Doctor, and only that agent runs in the sandbox on the local machine. Afterwards it can @mention a reviewer with the result.

Humans see the whole conversation, while an agent's context stays narrower because BAND delivers only messages that mention it. Agents that need their own isolated environments can run in separate sandboxes and still share a room.

What it means for developers

Docker's pitch is that developers avoid building agent identity, routing, a worker connection back to the laptop, credential handling and an audit trail before reaching the actual task. Author Oleg Selajev writes that the link between BAND and the sandbox is "a WebSocket connection and a room mention, which is about as much integration code as I would like this problem to have."

For verification, Docker suggests checking sbx policy log to confirm that the sandbox's network requests match the intended policy. The BAND Python SDK repository holds the SDK and kit sources, and the full Repo Doctor example sits in the Docker Samples repository. Kit configuration, credentials and network rules are covered in the Docker Sandboxes kit documentation.