The official BAND AI kit for Docker Sandboxes is now available at docker.io/bandhq/band-python-kit:latest. This kit packages all necessary components for an agent to operate within a Docker Sandbox, allowing a single `sbx create` command to yield a ready-to-use agent environment.
However, an agent confined to an isolated sandbox requires connectivity to perform meaningful tasks. The BAND AI kit facilitates this by linking a sandboxed agent to external systems. The agent operates on the host machine, gaining access to the local repository, and joins a BAND room to interact with both human participants and other agents, regardless of where they are running—locally, in the cloud, or on private infrastructure.
Furthermore, the kit enables communication between multiple sandboxes, permitting several agents to collaborate within a single room while each maintains its own segregated operational environment. BAND provides the necessary collaboration layer for these agents, allowing developers to grant persistent identities, convene agents in shared rooms, direct tasks using `@mentions`, track message history, and surface tool execution events within the room.
Each agent possesses its own infrastructure, framework, model, tools, and workspace, joining the room as a participant based on its execution location. Docker Sandboxes provide this isolation via a microVM, complete with its own kernel, filesystem, Docker daemon, and network. Within this environment, the agent interacts with local source files, executes commands, runs tests, and modifies a mounted workspace.
The BAND kit bridges this isolation and connectivity using one `sbx create` command. Connectivity is unidirectional; the agent initiates an outbound WebSocket connection to BAND and transmits replies via the BAND API, ensuring no ingress is required into the sandbox. From the room's perspective, the participant behaves like any other entity.
For instance, Repo Doctor, a small agent designed to repair a faulty Python project, runs inside a local Docker Sandbox and participates in a BAND room. Work is assigned to the agent by mentioning its handle via the BAND platform, irrespective of its physical location. The example repository includes a `band.yaml` file that directs the BAND sbx kit on which agent to launch, defining instructions, selecting Claude as the model, and provisioning tools for file reading, code editing, and testing.
To implement this demonstration, participants require Docker Sandboxes installed and authenticated, a BAND account, and an Anthropic API key with associated credits. The process involves setting custom secrets—the BAND agent key and the Anthropic API key—which Docker Sandboxes securely hold on the host, exposing only placeholder values to the sandbox.
The agent is then instantiated using the kit, allowing access to both BAND and the Python package hosts. After configuring network policies to permit access to `api.anthropic.com`, the agent is launched. Upon connecting to the platform, the agent can be added to a BAND room, where its handle allows other participants to issue commands.
This architecture supports complex compositions, where a room might contain hosted agents (like a Planner or Research agent) collaborating with a local agent, such as Repo Doctor, running in a Docker Sandbox. Only the local agent requires access to the repository workspace, while the entire conversation history remains visible to human participants.
This design significantly streamlines development by abstracting away the need for building agent identity, message routing, worker connection back to the host, credential management, and audit trails, as BAND handles this composition via a WebSocket connection and room mentions. This trend reflects an operator's need for secure, composable execution environments for AI workloads.
