AGENT DEVICE CLOUD
Updates and articles
Understand why ADC exists, how agents and devices work together, and what each release changes. Engineering articles explain design choices, guides provide steps, and release notes describe shipped behavior.
All published content15 entries
Android 0.2.8, CLI 0.1.9 and SDK 0.2.5: screenshot recovery and MCP imagesRecover a stalled screenshot stream within the existing sharing session and deliver screenshot pixels directly to MCP clients.Android 0.2.7, CLI 0.1.8 and SDK 0.2.4: steadier screen operationsReuse the connection, observe the result of each screen action, and recover from stale interface references without replaying uncertain actions.CLI 0.1.7 and SDK 0.2.3: invocation intent and control sessionsDescribe the purpose of a tool call, keep screen actions in their authorized control session, and install the current CLI and SDK from official downloads.Connector 0.1.6: teams, device sharing and root sharingShare a device with a team or a named account, choose how shared work runs, and see which devices need an update.When a cloud agent needs your computer: the device access layerPersonal agents can reason and use cloud tools, but your files, programs and hardware may live elsewhere. Why device access becomes necessary, how ADC supplies it, and what cross-platform Nodes make possible.Why separate an agent from the devices it works on?An agent can plan in one place and use tools on several devices. Where ADC fits, what it provides, and what remains the responsibility of the agent and device.Connector 0.1.3: bounded task waits and the official serviceFollow a long-running task without submitting it again, use the official service address, and choose the installation for your device.Connector 0.1.1: faster mobile control and visible updatesHow the release handles changing Android screens, keeps long-running MCP sessions current, and helps find outdated Connectors.From the first device to the first verified tool callPair one computer, expose a test folder, create read-only Agent access, and verify the result and task record.Least privilege for AI agents is an intersection, not a checkboxFollow a read, a denied write and an approved test to see how account access, Agent grants and device policy work together.Durable receipts help recover agent side effectsWhat to do when a command may have finished but its response was lost: follow the task, preserve its identity and reconcile uncertain effects.Faster task delivery with outbound WebSocket wakeupsWhy the Connector now receives a small wake signal, how tasks are still claimed durably, and what operators should check when delivery slows down.How an AI agent reaches a computer behind NATTrace a cloud request to a local file through outbound connections, durable task delivery and device-side checks, including sleep and network failure.Choosing between ADC, SSH and remote desktopThese approaches overlap. Compare what you need to connect, how commands are invoked and what the caller must manage.Continue local development from a remote AI agentAn example workflow for inspecting a real repository, making a focused change and running its configured test command on the existing computer.