Claude Code on the web, or a machine of your own: what each keeps

What a cloud session is

Anthropic’s docs now call it a cloud session, and the browser at claude.ai/code is one way in; the Code tab in the Claude app, the desktop app with Cloud selected, claude --cloud from a terminal, and routines, which run each scheduled run as a cloud session, are others. By default it runs on Anthropic’s machines; Team and Enterprise organisations can route sessions to a self-hosted environment on their own compute, in public beta and off by default. In Anthropic-hosted environments, in Anthropic’s words, “each session gets a fresh virtual machine (VM) running Ubuntu 24.04 on x86_64, regardless of your own operating system and CPU architecture, with your repository cloned and common toolchains pre-installed.” The environment page gives the machine as about 4 vCPUs, 16 GB of memory and 30 GB of disk, “approximate resource ceilings that may change over time,” and lists what is installed: Python, Node, Ruby, PHP, Java, Go, Rust, C and C++, Docker, PostgreSQL 16, Redis 7, git, the GitHub CLI and tmux. A setup script you add runs as root, so apt install works, and the environment keeps what it installs (below).

The code comes from GitHub. A session can clone any public repository; a private one needs the Claude GitHub App installed on it, or a token you hand over with /web-setup (on Team and Enterprise, once an Owner has turned Quick setup on). From the terminal, the session clones your repository’s remote at your current branch, “not your local checkout, so push first if you have local commits.” Run claude --cloud in a repository with no remote, or a github.com repository without the App, and Claude Code bundles the repository and uploads it instead: its history, tracked files and uncommitted changes to them, with fallbacks to the current branch and then a snapshot above about 100 MB. A GitLab or Bitbucket repository goes up only when you set CCR_FORCE_BUNDLE=1, and then “the session can’t push results back to that remote.” You can add several repositories to one session.

What survives a session, and what doesn’t

• The branch. When Claude reaches a stopping point it pushes a branch. You review the diff in the browser, leave comments, open a pull request from there (full, draft, or GitHub’s own compose page), or send another message. The session doesn’t close when the branch is pushed, and each commit carries a Claude-Session trailer pointing back at it.
• The conversation. Anthropic stores the transcript so you can return to the session. If its machine has been reclaimed in the meantime, what comes back is “your conversation history,” and what does not is “background work that was still running when the VM was reclaimed, such as subagents and shell commands.”
• The environment, as a snapshot. If your setup script finishes within about five minutes, Anthropic snapshots the filesystem and starts later sessions from it, skipping the script; the snapshot rebuilds on its own when you change the script or the allowed hosts, and after roughly seven days. It “keeps what the setup script writes to disk and loses anything that was only running”: a database the script started or a docker compose up stack is not restarted for you, so you ask Claude to start it or add a SessionStart hook. Packages Claude installs during a session “don’t carry over to other sessions”; packages the setup script installs do.
• Your own configuration, by two routes. A committed CLAUDE.md, rules, skills, agents and commands load; a repository’s hooks and project .mcp.json load in a session with one repository; plugins a repository declares don’t install. Your user-level ~/.claude/CLAUDE.md, user skills and plugins, and MCP servers you added at local or user scope do not come along: “To make your own configuration available in cloud sessions, commit it to the repo,” or let the setup script write a ~/.claude/CLAUDE.md for what you would rather not commit. Skills you enable on claude.ai load on their own.

The idle stop

This is the behaviour that decides whether a cloud session fits your work. Closing the tab doesn’t stop anything: the session “continues running in the background until Claude finishes the current task, then idles.” Then, in two stages: “after a few minutes without activity, a session’s VM pauses with its files saved. Your next message restores the same VM,” and later, “if the paused VM has since been reclaimed, reopening the session provisions a fresh VM.” You can’t pause or hold a machine yourself, and the docs publish no figure for how long a paused machine is kept. If Claude asks you a question and you are away, you can still answer when you come back, “up to environment expiry.”

The command limits are the Bash tool’s own, the same as on your machine: a foreground command gets two minutes by default and can ask for ten, and a background command gets 30 minutes, or up to two hours when Claude asks for it, with the environment’s variables (BASH_DEFAULT_TIMEOUT_MS, BASH_MAX_TIMEOUT_MS) raising both. What a cloud session has no setting for is a process that outlives the session: it is a worker that runs tasks and stops, not a machine that stays up.

What it costs, and who has it

“Cloud sessions are available on Pro, Max, and Team plans, and for Enterprise users with premium seats or Chat + Claude Code seats”; on newer Enterprise plans Claude Code comes with the single seat, and whether cloud sessions do is not stated. On Team and Enterprise an Owner turns them on (the allow_remote_sessions policy). They “share rate limits with all other Claude and Claude Code usage within your account,” and “there is no separate compute charge for the cloud VM.” Several sessions can run at once, each started by its own --cloud, and “running multiple tasks in parallel consumes more rate limits proportionately.” It needs an Anthropic account: not available through Amazon Bedrock or another provider, not for organisations with zero data retention or the HIPAA configuration, and pulling a session down to a terminal needs the claude.ai sign-in rather than an API key.

What you can’t do there

“You don’t get a shell into the session VM. Claude runs every command for you.” The docs describe only outbound traffic, all of it through Anthropic’s proxy, and no way to reach a service running inside the machine; combined with the idle stop, that rules out a dev server, a database or a bot that stays up. Network access is “Trusted” by default, an allowlist of package registries, GitHub and cloud SDKs, with None, Full and a custom list as the alternatives; Claude’s own calls to the Anthropic API go through even at None. GitHub access goes through Anthropic’s proxy, so your GitHub credentials “never enter a session’s VM” unless you set a token on the environment yourself, which passes through to it. Terminal-only commands such as /plugin and /resume aren’t available, nor are the Manual and Bypass permission modes, and a sign-in that needs a browser (SSO) can’t run there. Built-in tools read issues and pull requests and post comments; the GitHub CLI’s gh pr and gh issue get a 403 because they use GraphQL, and gh api covers the rest.

Getting the work into a terminal

claude --teleport <session-id> “fetches and checks out the branch from the cloud session, and loads the full conversation history into your terminal.” It wants a clean working directory in a checkout of the same repository, the branch pushed, and the same claude.ai account. From then on the terminal has its own copy; new work there stays local. The handoff is one-way from the CLI: you can pull a cloud session down, but “you can’t push an existing terminal session to the cloud.” Only the desktop app sends a local session up, “with your conversation carried over as a summary.”

When a machine of your own is the better shape

Anthropic draws the line itself: “For work that needs your local config, tools, or environment, running Claude Code locally or using Remote Control is a better fit.” And on when each applies: “Use Remote Control when you’re in the middle of local work and want to keep going from another device. Use a cloud session when you want to start a task without any local setup, work on a repo you don’t have cloned, or run multiple tasks in parallel.” Read against the list above, a machine of your own wins when:

• something has to keep running: a dev server, a database, a browser signed in to the sites the work touches, a bot, a process that takes hours;
• the tools and configuration you installed once should simply be there, with no setup script to maintain and nothing to commit for the sake of the machine;
• the code isn’t on GitHub, or a push back to its home matters;
• the job needs more than about 4 cores and 16 GB (the docs say nothing of GPUs); Anthropic’s other answer for that, on Team and Enterprise, is a self-hosted environment;
• you want the session itself, not a branch: open it, type into it, leave it for a week and find it where you left it.

The catch with your own computer is the one the phone guide spells out: Remote Control is a window into a session on your computer, so “your computer has to stay on and the claude process has to keep running.” A laptop that sleeps reconnects when it wakes; nothing ran while it slept. So the shapes that match a cloud session’s keep-going property are a desktop that never sleeps, a server, or a cloud computer of your own, with Claude Code started inside tmux so it survives your SSH connection dropping (server mode gives up after about ten minutes without a network and has to be started again), which is how Claude Code on a VPS is set up and what an orchestrator session needs. Everpod’s developer pod is that machine ready made: a cloud computer that is yours, always on, with Claude Code, Codex or both installed, reached only through your own private network, from $24 a month. A developer pod is one machine that stays yours: what you installed, the servers you left running, Docker and your work in progress are where you left them tomorrow.

Side by side

As of October 2026, from Anthropic’s docs and the facts above:

Where your code goes

In a cloud session “your repository is cloned into the session’s isolated VM,” the transcript is stored so you can return to it, and “code and session data are subject to the retention and usage policies for your account type”: on a Pro or Max account Anthropic trains on the data when that account’s training setting is on, and under commercial terms it does not. You can delete a cloud session, which removes its event data. On your own machine the files stay on it, but what Claude reads of them and what it answers still go to Anthropic’s API to do the work, under the same retention and training rules, and while Remote Control is connected “the session transcript, including your messages, Claude’s responses, and tool activity, is stored on Anthropic servers” too. The data-usage page carries the retention periods.