Remote ArcREMOTE ARC
Remote Arc Remote MCP reference
Reference for the Remote Arc MCP endpoint, OAuth authorization, device discovery, file and process tools, browser capabilities, durable task controls and client integration.
Endpoint and authorization
Compatible clients connect to https://mcp.remotearc.app/mcp. Remote Arc uses OAuth authorization for AI clients and keeps that authorization separate from device pairing. A client grant can be revoked without re-pairing every computer, and a single computer can be revoked without invalidating every AI connection on the account.
Device discovery and selection
The MCP surface can list computers linked to the account and report whether each device is online. Tool calls target a specific device identifier so work is routed to one selected connection rather than broadcast across the account. Device metadata is used for routing and policy decisions; the remote client does not receive an operating-system login session.
Files, processes and commands
Depending on device policy, MCP tools can list directories, read text or bounded binary data, inspect file metadata, inspect processes, apply supported text edits and start managed terminal commands. Write operations can be limited by Trusted Write Locations and protected-path rules. Terminal access is intentionally distinct because once enabled it inherits the local user's authority.
Browser companion capabilities
On a Chrome tab you shared, Remote Arc can expose a simplified snapshot, selected text, links or table data; click and fill need additional per-tab permission. On sites you allowed in the extension, shared_tab_open opens tabs and shared_tab_act clicks, fills, selects, presses keys, navigates and waits, with final submits sent once. Password, payment-card and file-upload fields are never filled.
Durable tasks and reasoning boundaries
Remote Arc can persist long commands, schedules, condition watches and bounded goal state so deterministic work can survive disconnects and chat endings. Task state and recovery do not imply that an AI model is continuously running. Any step that requires fresh interpretation still needs a reasoning controller to resume from saved context and choose the next allowed action.