What \"Claude Remote Agents\" Actually Means: Cloud Sessions, a Steered Local Session, or a Machine You Own

Every month, a few hundred people type "claude remote agents" or "claude code remote agent" into Google and land on our page about driving Claude Code from your phone. Almost none of them click, and they are right not to: that page answers one question, and the search asks three.

"Remote" is doing a lot of work in that query. Depending on who types it, it means an agent that runs on Anthropic's servers while the laptop is closed, an agent that runs on the laptop but is steered from a phone, or an agent that runs on a build server the person owns. Those are three different products with three different sets of rules, and Anthropic itself renamed the first one this year, which is where a good part of the confusion comes from.

This article separates the three, with the facts checked against Anthropic's documentation on 28 September 2026, and then says what we do for each case in AgentsRoom, with any agent CLI.

The word "remote" covers three different things

1. Cloud sessions: the agent runs on cloud infrastructure, not on your machine. Anthropic's former name for the CLI flag was --remote, which is why the search term stuck.
2. Remote Control: the agent runs on your machine, and you steer it from claude.ai or the Claude app on your phone.
3. A machine you own: the agent runs on a server, an office desktop or a home machine, and you reach it over SSH or a tunnel.

Two of the three are Claude Code features. The third is plain infrastructure, and it is the only one that works with every agent CLI.

1. Cloud sessions: the agent runs on Anthropic's machines

Anthropic's documentation calls it a cloud session: "a Claude Code session that runs on cloud infrastructure instead of on your machine". By default that infrastructure is an isolated VM Anthropic manages; an organization can route sessions to its own self-hosted environment instead. The session keeps running after you close your laptop, and you can check on it or steer it from any device.

You can start one from five places, and they all share the same cloud environments (network access rules, environment variables, setup script):

The terminal form is the one that explains the search term. The older spelling, --remote, still works as a deprecated alias of --cloud. The organization policy that allows or forbids the whole feature is still called allow_remote_sessions. So "remote" was the official word for a while, and it survives in flags, error messages and admin settings.

What actually happens when you run claude --cloud: a new session is created on claude.ai, and the cloud VM clones your repository's GitHub remote at your current branch. It clones the remote, not your working tree, so you push first if you have local commits. If the repository has no remote, or the Claude GitHub App is not installed on it, Claude Code bundles the local repository (under 100 MB, tracked files only) and uploads it, leaving out uncommitted .env, *.tfvars and key files. The task runs in the cloud while you keep working locally. When it is done, you create a pull request from claude.ai/code, or you pull the session back into your terminal:

claude --teleport                 # session picker
claude --teleport <session-id>    # a specific session

Teleport checks four things: a clean working directory, a checkout of the same repository, the session's branch pushed to the remote, and the same claude.ai account. It then fetches the branch and loads the whole conversation into your terminal. From the CLI the handoff is one way: you can pull a cloud session down, but you cannot push an existing terminal session up. Only the desktop app has a Continue in menu that sends a local session to the cloud.

Routines are the scheduled form of the same thing: a saved prompt, one or more repositories, a set of connectors, run on a schedule, on an HTTP call or on a GitHub event, each run as a full cloud session with no permission prompts. They are in research preview, created at claude.ai/code/routines or with /schedule in the CLI, and they add a daily cap on runs per account.

Three limits worth knowing before you rely on it. Cloud sessions need GitHub: cloning and pull requests go through the Claude GitHub App or through your gh token sent with /web-setup. They share rate limits with everything else on your account, with no separate compute charge but proportional consumption when you run several in parallel. And they require a claude.ai login on Pro, Max or Team (Enterprise with the right seats): --cloud and --teleport do not work with an API key, nor when Claude Code is configured for Bedrock or Vertex.

2. Remote Control: the agent runs on your machine, you steer it from elsewhere

Remote Control is the mirror image. The documentation states it in one sentence: "Unlike cloud sessions, which run on cloud infrastructure, Remote Control sessions run directly on your machine and interact with your local filesystem." claude.ai/code and the Claude app on iOS and Android are a window into a process that is still running on your laptop, with your files, your MCP servers, your hooks and your project configuration. Typing @ on the phone even autocompletes paths from your local project.

Three ways to start it:

claude remote-control            # server mode: serves sessions to your devices
claude --remote-control          # an interactive session you can also drive remotely
/remote-control                  # from inside a session that is already running

The first time, Claude Code asks you to confirm once. It needs a claude.ai login on Pro, Max, Team or Enterprise (API keys are not supported), and on Team and Enterprise an owner has to enable the toggle first. If the laptop sleeps or the network drops, the session reconnects when the machine comes back. And the note in Anthropic's own documentation is the one to remember: --cloud creates cloud sessions, --remote-control is unrelated.

The two features do meet in one place. --teleport travels over the Remote Control session infrastructure, so an expired teleport shows errors with Remote Control wording. That is a plumbing detail, not a sign that the two are the same thing.

If the question you have is about tokens, we measured it: Remote Control does not consume more tokens, and here is what Codex has instead.

3. A machine you own: SSH, a tunnel, a build server

The third meaning has nothing to do with Anthropic's account. The agent runs on a computer you own that is not the one in front of you: a build server, the office desktop, a VPS. You reach it over SSH, you run claude or codex there, and the session lives on that machine.

Editors are starting to package this. VS Code 1.121 added "Remote agents (Preview)": the Agents window can run agent sessions on a remote machine you own, over SSH or dev tunnels, through a long-lived agent host process, so a running session continues even if your client disconnects. Same idea, one CLI, one editor.

The plain version is a terminal and ssh. It works with every CLI, it keeps your code on hardware you control, and it has the usual costs: the session dies with the connection unless you wrap it in tmux or screen, and the phone is not a great SSH client.

How to choose

The right question is not "which one is best", it is where you want the code to run and what you need to keep.

Two things to keep in mind. Cloud sessions and Remote Control are Claude Code features: there is no Codex equivalent of --teleport, and Antigravity CLI ships its own Remote Control since version 2.0. And "remote" does not mean "unattended": a cloud session that asks a question waits for you until its environment expires, and a Remote Control session waits for a permission answer like any local one.

What we do for each case

AgentsRoom runs the agents, it does not host them, so it takes a side on each of the three cases rather than replacing them.

Steering a local agent from the phone, for any CLI: the AgentsRoom mobile app takes over the terminal of every agent running on your desktop, Claude Code, Codex, GitHub Copilot CLI, Cursor and the other agent CLIs, over an end-to-end encrypted relay, with push notifications when an agent waits for you. It coexists with Anthropic's Remote Control on a Claude Code session: one is a window into the conversation, the other into the terminal.

An agent on a machine you own: SSH connections save the server once and run the CLI there, with the same tiles, statuses and notifications as a local agent; Remote Fleet does the same across every computer of your account, office, home and build server in one view.

No machine of yours at all: Cloud Agents start a disposable machine that clones the repository, runs Claude Code or Codex headless on the instruction you sent from your phone, commits, pushes a branch and destroys itself. It is a job, not a session: there is no screen to connect to, which is exactly the difference with the two cases above.

None of this changes what Anthropic's cloud sessions do. If you are on Claude Code and you want Anthropic to run the VM, claude --cloud is the tool, and the branch it pushes shows up in your repository like any other.

Frequently asked questions

Are Claude remote agents the same thing as Claude Code on the web?

Yes. Anthropic's current name for it is a cloud session: a Claude Code session that runs on infrastructure Anthropic manages (or on your organization's self-hosted environment) instead of on your machine. You start one from claude.ai/code, from the Code tab of the Claude app, from the desktop app by picking Cloud instead of Local, or from the terminal with claude --cloud. The word remote survives from the older flag, --remote, which still works as a deprecated alias of --cloud, and from the organization policy that governs it, allow_remote_sessions.

Does claude --cloud run anything on my computer?

No. It creates a new session on claude.ai and the cloud VM clones your repository's GitHub remote at your current branch, not your local checkout, so you push first if you have local commits. If the repository has no remote, or the Claude GitHub App is not installed on it, Claude Code bundles the local repository and uploads it instead, leaving out uncommitted .env, key and credential files. The task then runs in the cloud while your terminal is free.

Does Remote Control run in the cloud?

No. Remote Control does the opposite of a cloud session: the Claude Code process keeps running on your own machine, with your filesystem, your MCP servers and your project configuration, and claude.ai/code or the Claude app on iOS and Android become a window into it. You start it with claude remote-control (server mode), claude --remote-control in an interactive session, or /remote-control inside a running one. Anthropic's documentation says in as many words that --cloud and --remote-control are unrelated.

Can I use Claude's cloud sessions or Remote Control with Codex, Antigravity or another CLI?

No. Both are features of Claude Code, tied to a claude.ai subscription. Antigravity CLI has its own Remote Control since version 2.0. For the other CLIs, the piece you can add is the one that does not depend on the vendor: AgentsRoom's mobile app takes over the terminal of any agent running on your desktop, whatever its CLI, through an end-to-end encrypted relay, and it coexists with Claude's Remote Control on a Claude Code session.

Do cloud sessions cost extra?

There is no separate compute charge for the cloud VM, but cloud sessions share rate limits with all other Claude and Claude Code usage on your account, and several sessions in parallel consume them proportionately. They are available on Pro, Max and Team plans, and on Enterprise with premium seats or Chat + Claude Code seats. Routines, which run as cloud sessions, add a daily cap on the number of runs per account.

How do I bring a cloud session back into my terminal?

With claude --teleport, which opens a session picker, or claude --teleport <session-id> for a specific one; /teleport does the same from inside a running session. Claude Code checks that your working directory is clean, that you are in a checkout of the same repository and signed in to the same account, then fetches the session's branch and loads the conversation history. The terminal gets its own copy: what you do next stays local and does not appear in the cloud session. The handoff is one way from the CLI; only the desktop app can send a local session to the cloud.