openclaw-dev is a Microsoft-published template for running your own always-on OpenClaw assistant in the cloud rather than on your laptop. The rationale is safety: OpenClaw executes arbitrary code and can be deceived by prompt injection, so the template hosts it in a disposable Azure Container Apps container instead of on a machine that holds your credentials. The agent talks to Azure OpenAI in Foundry Models through a Managed Identity — local auth on the model endpoint is disabled, so no API keys exist anywhere in the system — and an optional sandbox execution mode offloads every untrusted tool run (shell, code, browser) to an ephemeral, isolated ACA Sandbox that is destroyed after the task.
Deployment is driven by devclaw, a thin wrapper over the Azure Developer CLI, with commands to deploy, stop (scale to zero at no cost), stream logs, switch execution mode, and tear everything down. The repo also ships an agent skill in the open Agent Skills format, so a coding assistant such as Copilot Chat or Claude Code can perform the setup, troubleshooting, and teardown from plain-English requests, following the repo's own scripts and error catalog. An optional Microsoft Teams channel — via an Azure Bot registration and the msteams OpenClaw plugin — makes the assistant reachable from your phone; a WebChat UI behind Entra ID Easy Auth is the default interface.
Security is layered: a Microsoft login gate in front of the container, a per-container gateway token for the WebSocket API, keyless managed-identity auth to the model, and the ephemeral container itself, with state persisted on Azure Files so a full nuke-and-redeploy takes about six minutes. It is explicitly alpha, with no production-readiness guarantees, and the README is candid about remaining risks: skills run arbitrary code, prompt injection remains possible, and the container runs as root by default.
