Which Remote Runtime Can Build Your Checkout?

Use VS Code Remote SSH when the checkout and toolchain already live on your host. Compare Cursor-hosted Cloud Agents and Cursor Self-Hosted Machines by testing build prerequisites and checking where the agent loop runs.

Test the environment from the agent’s execution context

VS Code Remote SSH runs commands and extensions on the remote machine, where you can open the host’s existing folder. That makes it a practical option for a repository that depends on an installed compiler, a particular CPU architecture, or private build services. Create a read-only preflight for a disposable checkout at a known revision. Record the revision, architecture, toolchain version, and whether the same account that runs the agent can read approved build inputs. A shell opened under a different account can give a false pass. For example, a package cache visible in your login shell may be missing from the agent’s execution context. Run the same preflight in each candidate environment and record the first failed prerequisite before timing any builds.

Remote Development using SSH

Check tool location and agent-loop policy before benchmarking

Cursor documents two execution choices for Cloud Agents. Its hosted VMs use x86_64, while a self-hosted worker can run on arm64 hardware. With Self-Hosted Machines, tool calls execute on your hardware and the agent loop stays in Cursor’s cloud. Consider an illustrative arm64 build that also reads an internal registry. A hosted VM could pass registry reachability and still fail a native arm64 requirement. A self-hosted worker could satisfy both runtime prerequisites and still be unsuitable if your policy prohibits a Cursor-hosted agent loop. Record policy decisions independently of preflight results. If you evaluate Remote SSH, review the coding agent you plan to launch there for its own model and data handling. An SSH connection establishes where the editor’s remote commands execute.

Choose where Cloud Agents run | Cursor Docs

Keep the checkout on your host

Use Codex CLI on your own always-on machine when the active checkout or build dependencies already live there. Codex CLI runs on the computer where you launch it. Attach to that machine from a second computer over authenticated SSH and confirm the repository path and branch in the remote shell before you start the agent.

For the DIY option, use Tailscale for private reachability and a named tmux session for the agent terminal. SSH is the baseline connection. If you prefer Mosh for a mobile link, test it on the actual client and network you plan to use. Close the client, reconnect, and reattach. A fresh agent prompt or a missing task transcript means you still need to prove session continuity.

GitHub - openai/codex: Lightweight coding agent that runs in your terminal · GitHub

Rehearse a handoff from a second device

VS Code Remote - SSH can open the remote folder and run commands on the host. Use it when the operator needs to review a diff in an editor. For phone control, evaluate Claude Code Remote Control for a Claude Code session, a Telegram or Discord bridge for chat, or a QR-paired terminal relay for terminal interaction. Verify each candidate's supported host, authentication and approval behavior before pairing a real repository.

In an illustrative trial, start a task that adds one line to a README on a scratch branch, then disconnect the originating client for five minutes. From the second device, send a follow-up that changes the requested wording. Pass the trial only if the original run receives it and the expected one-line diff appears in the remote checkout.

Remote Development using SSH

Tell a dropped link from a stopped run

Cursor offers a useful distinction for teams considering a managed control surface with local execution. Cursor-hosted Cloud Agents run in managed VMs, while Self-Hosted Machines execute tool calls on hardware you control and leave the agent loop in Cursor's cloud. Use the latter when tool execution must stay on your hardware and your policy permits the cloud agent loop. Confirm where tool results may go before using an internal repository.

For any pilot, record the host, session identifier, scratch branch and expected output before disconnecting. If the phone stops updating, check the task from an independent operator session. Advancing output points to a control-link problem. A missing process or a worker that cannot execute a simple read points to the host or worker. Inspect the last log entry and diff before starting another run.

Choose where Cloud Agents run | Cursor Docs

Sources