REMOTE ARC
Remote MCP Server: what it is and how to connect
Remote MCP lets an AI client discover and call tools exposed by a server reachable over the network. Learn how this differs from locally launched MCP processes, how to connect a compatible ChatGPT, Claude or Cursor client, and how Remote Arc uses one hosted MCP endpoint to bridge approved tools on computers you explicitly pair.
What is a remote MCP server?
MCP, the Model Context Protocol, defines how a client discovers and invokes tools made available by a server. With a local MCP setup the AI host commonly starts the tool server as a child process and exchanges protocol messages over standard input and output (stdio). A remote MCP server is reachable through a network transport such as Streamable HTTP, so the AI host need not launch that MCP server as a local child process. Authentication and individual client compatibility still matter; a reachable URL alone does not make every MCP client able to connect or use every tool.
Local MCP versus remote MCP
Local stdio MCP is useful when an AI client and tools live in the same development environment, or when you can configure and run a tool process next to your assistant. Remote MCP is useful for managed network services, centrally controlled OAuth integrations and clients that need tools hosted elsewhere. The word remote describes how the client reaches an MCP endpoint. It does not require every downstream operation to run on the server hosting that endpoint. Remote Arc, for example, routes permitted calls from a hosted MCP control plane to the local agent on a paired computer.
How to connect a Remote MCP server in three steps
First, choose a supported remote MCP server and understand what capabilities it provides. For Remote Arc, run npx remotelink on the Windows, macOS or Linux computer that should receive tool calls, then approve pairing through the Remote Arc dashboard. Second, open the MCP apps or connectors configuration inside your AI client and add https://mcp.remotearc.app/mcp using its supported remote HTTP setup. Complete OAuth to authorize that client to your Remote Arc account. The menus and the availability of custom remote MCP connections depend on client version, account and workspace settings. Third, list paired devices through the connected client and make a harmless read-only test call before authorizing file edits, shell commands or other broader tools.
Remote MCP setup for ChatGPT, Claude and Cursor
ChatGPT requires a supported app or MCP connector entry point; Remote Arc can be installed when available or connected through the appropriate developer setup. Claude exposes custom connector setup with OAuth where supported. Cursor supports remote MCP server configuration; confirm authorization and tool discovery before requesting work on a project. Follow the current Remote Arc install guide for each client instead of copying a configuration meant for a different application. Running a computer agent alone does not automatically add its hosted MCP endpoint to your ChatGPT, Claude or Cursor account.
Remote Arc architecture: from AI client to your computer
Remote Arc separates the AI's reasoning from tool execution. Your AI client sends authenticated MCP requests over HTTPS to the Remote Arc endpoint. The hosted service checks account authorization and the selected device's policy, then routes the request through an outbound connection already established by that computer's local agent. Files, processes and approved terminal commands run on the selected computer. Ordinary setup does not open an inbound public MCP port on your laptop, expose a router port or require a VPN. A device still needs power, network connectivity and a working local agent to receive calls.
Security: OAuth is not permission to run everything
OAuth grants a supported AI client access to the Remote Arc account, not a shared operating-system login or unconditional control of every paired machine. Each computer has its own revocable device identity and configurable tool permissions. Start with read access; enable writing or terminal calls only for workflows you trust. Trusted Write Locations and sensitive path rules narrow supported file operations. Terminal commands, when separately enabled, run under the local OS user and are not placed in an operating-system sandbox by those file rules. Requested data and tool results can pass through the hosted relay and AI provider, so review the security and privacy documentation.
Troubleshooting remote MCP connections
If a client cannot discover tools, verify that your particular AI account supports remote MCP apps or connectors, that the endpoint is configured in that client and that OAuth completed. If the client connects but a computer appears offline, inspect the computer's power and network state and run npx remotelink to confirm the local agent is available. If a tool is denied, check the exact target device and its permissions; a successful OAuth login is separate from write or terminal authorization. The Free plan includes background agent recovery, but planned durable jobs, scheduling and keep-awake are separate Plus capabilities. Neither tier can execute on a powered-off computer.