Claude Code on the web runs a cloud environment on Anthropic's infrastructure: it clones a GitHub repository and lets you review, commit, push and PR within the granted GitHub authorization and branch permissions. The key difference from local work - this environment has no local files or shell state of yours; dependencies and setup are determined by the repository and the environment configuration. Tasks can be passed between the terminal and the web by official commands, but prompt injection from the repository stays just as relevant here.
Remote Control is a fundamentally different model, and it must not be confused with the web. It does not move computation to the cloud: Claude Code keeps running on your machine, and claude.ai or the mobile app becomes a remote client. A suitable claude.ai subscription and a full login are needed; API-key authentication is not supported. Files stay on the host, but any confirmed remote prompt runs local tools - so such a session must not be left on an unlocked machine with production credentials.
Remote Control is launched from the shell or from an already open session. It helps to see the commands once: an interactive local session available remotely, and a server mode for connecting sessions, and from within - /remote-control or /rc. The connection goes over an outbound channel. Server mode can create sessions in the same directory, a separate worktree or single-session mode, and eligibility and flags depend on the account and version - so you check the current --help rather than memory.
The mobile app is a window into cloud and Remote Control sessions, not a full environment. It shows such sessions, lets you send prompts and answer permission requests, and push notifications are handy for completion and approval. Mode limits differ: cloud sessions give no bypass, and Remote Control reflects the actual local flow with documented display exceptions. Mobile is observation and answers on the go, not a replacement for the local machine.
Chrome and computer use extend the surface to the browser and GUI - and at the same time the risk surface. The Chrome integration gives navigation, forms, console, screenshots, downloads and multi-site workflows via an extension; site permissions are granted explicitly, and a web page is untrusted input. Computer use drives approved apps via screenshots and actions; one session is available at a time, the work can always be stopped. In a mixed-trust task you must not allow access to a password manager, a bank, a production console or a personal messenger.
Artifacts is a shareable interactive page from the session output: a walkthrough, a comparison, a configurable demo, a report. It is important to understand its boundaries: it is not a repository deployment and not a guaranteed long-term storage. Viewer-side connector calls may require their own authorization, and public sharing, connectors, retention and audit are governed by organization settings. An artifact is handy to show a result, but not to keep the only copy of something important in it.
There are three schedules, and each has its niche. It helps to gather them into a table once: a cloud Routine runs in Anthropic's cloud, needs no host, a minimum interval of an hour, durable. A Desktop scheduled task runs on your machine (on, but without an obligatorily open session), an interval from a minute, durable. /loop and cron tools live in the current CLI session, need it alive and are durable only until resume or expiry. /loop has forms: with an interval, with a self-chosen delay and bare for a maintenance prompt.
| Mechanism | Where | Needs host/session | Min interval | Durable |
|---|---|---|---|---|
| cloud Routine | Anthropic cloud | No | 1 hour | Yes |
| Desktop scheduled task | Your machine | Machine on | 1 minute | Yes |
| /loop / Cron | Current CLI session | Session must live | 1 minute | Until resume/expiry |
Choosing a scheduler is a decision about dependencies and reliability. Cloud Routines are created via the web or /schedule, they have schedule, API and GitHub triggers, a fresh environment and autonomous runs. Desktop tasks are needed when local files and tools are required. And /loop is not chosen for a critical unattended SLA: it is tied to a live session, tasks expire, and it is a handy but not guaranteed mechanism. The typical failure is assigning /loop what must run unattended; the right move is a Routine or a Desktop task with an owner, a deadline and audit.
# Remote Control: computation stays on your machine
claude --remote-control "Release check" # interactive, available remotely
claude remote-control # server mode to connect sessions
# from an open session: /remote-control or /rc; check --help (eligibility)