Claude / Operating Manual · Reference

← All Operating Manual pages

Remote & Web

Run and steer Claude Code beyond one terminal: remote sessions, cross-device Remote Control, and the desktop GUI.

preview · updated 2026-08-24

Claude Code increasingly runs beyond a single local terminal. Remote Control (previewed at Code with Claude, 2026-05-06) lets you start a session on one device and continue it on another; it shows as a persistent footer pill (2.1.162, 2026-06-03). The IDE extensions also added remote agent sessions that run the agent on a remote machine over SSH or a dev tunnel. The desktop app got a redesign with split views, pinnable assistant messages as chapters with an auto-generated table of contents, and inline diff commenting. Earlier, the Cowork desktop app wrapped the same agentic primitives (filesystem, MCP connectors, skills) in a point-and-click UI.

The terminal CLI is still the canonical, most complete surface. These are ways to reach it from elsewhere, not replacements.

How you use it day-to-day

  • Cross-device Remote Control. Kick off a session on your laptop, pick it up on another device. It needs login auth (see the gotcha below).
  • Remote agent sessions. Point the agent at a remote box over SSH or a dev tunnel when the work belongs where the code/compute lives: big repos, cloud dev envs.
  • Desktop GUI for review-heavy work. Split views, chaptered messages, and inline diff comments are the win over the terminal when you're reviewing a lot of output.
  • Self-hosted runners. claude self-hosted-runner (2.1.224, Team and Enterprise) turns your own machines or containers into a place web, mobile, and desktop sessions can run, so remote sessions execute next to your code and secrets instead of in Anthropic's cloud.
  • Resume the last Remote Control session. claude remote-control --continue (2.1.229) picks up the most recent one without hunting for it.

Gotchas

  • Remote Control needs login auth. It's disabled when ANTHROPIC_API_KEY, an auth token, or apiKeyHelper is set (2.1.139). It keys off your logged-in account, not an API key.
  • These surfaces are young and moving fast (preview). Check the changelog before depending on a specific capability.
  • Connector reliability was fragile in early desktop previews. Verify an MCP connector is actually live before building a task on it rather than looping on retries.

The web surface (claude.ai/code) exists but isn't documented in detail here yet. See the official docs for its current capabilities.