Claude Code on the web is the cloud version of Claude Code: at claude.ai/code you start a Claude Code session in your browser that does not run on your computer but on a virtual machine that Anthropic runs by default. You connect GitHub once. Claude clones your repository into the VM, works there, runs tests and pushes a branch back, which you review in the browser and turn into a pull request. The session keeps running when you close your laptop, and you can check in from your phone or the desktop app. According to Anthropic, cloud sessions are available on Pro, Max and Team, and for Enterprise users with a premium seat or a Chat + Claude Code seat. There is no separate charge for the cloud VM; sessions count toward your plan's usage limits. From the terminal, you start a cloud session with claude --cloud and pull it back with claude --teleport. Scheduled tasks, formerly "routines", also run as cloud sessions.
Five of my AI employees run on their own Linux server and are steered through Remote Control, as described in Claude Code on a Server: tmux, Remote Control. This article explains the cloud version based on Anthropic's docs as of October 26, 2026; I am not describing a hands-on test of cloud sessions of my own here. Everything about Claude is collected on my Claude page.
What "Claude Code web" means, and what it doesn't
Three things get mixed up under this term. Usually it means the cloud session, which Anthropic itself calls "Claude Code on the web". Then there is Remote Control: you also open it at claude.ai/code, but the session runs on your own computer, and the browser is just the window. And some people simply mean building a website with Claude Code; more on that in the FAQ below.
The docs compare the three ways to run Claude Code like this:
I explain how Remote Control works in detail in Claude Code Remote Control from Your Phone.
What cloud sessions are good for
Anthropic names four cases: several independent tasks in parallel, each in its own session and on its own branch, without managing worktrees yourself; repositories you don't have locally; well-defined tasks that need little steering; and questions about a codebase without checking it out. For work that needs your local configuration, tools or environment, the docs recommend a local session or Remote Control instead.
A typical example: you have three small, independent tasks in a GitHub repository, say a login bug, missing tests and an outdated dependency. In the cloud you start three sessions for them, each on its own branch, and review three diffs in the evening. Locally you would have to keep three terminals or three worktrees open.
Your first cloud session in the browser
- Open claude.ai/code and sign in with your Claude account.
- Connect GitHub. Sessions can then reach all public repositories, but private ones only if the Claude GitHub App is installed on them. On Team and Enterprise, an Owner first has to turn on the organization's GitHub connector.
- Set up an environment. On Pro and Max, onboarding creates an environment called "Default" for you; on Team and Enterprise you fill in a short form. "Default" has "Trusted" network access: package registries, GitHub and other allowlisted domains, nothing else.
- Describe the task. Claude clones the repository, runs your setup script if there is one, does the work and runs tests.
- Review the result. Claude pushes a branch to GitHub. In the browser you see the changes as a diff, comment on individual lines and create a pull request. The session stays open afterward, so you can keep giving instructions.
If you already use the GitHub CLI gh, you can also connect GitHub from the terminal with /web-setup. On Team and Enterprise this only works once an Owner turns on the organization setting "Quick setup".
From the terminal: --cloud and --teleport
You can start a cloud session straight from the terminal. The example from the docs:
claude --cloud "Fix the authentication bug in src/auth/login.ts"
The cloud VM clones the GitHub state of your current branch, not your local checkout. So push local commits first. Without a GitHub remote, or if the GitHub App is missing on the repository, Claude Code uploads your local repository as a bundle. According to the docs, on macOS, Linux and WSL it leaves uncommitted changes to files such as .env or key files on your machine. The old spelling --remote still works but is deprecated.
You pull a cloud session back into the terminal with:
claude --teleport
Without an argument you get a picker; with a session ID you jump straight in. Claude checks that you are in the right repository, fetches the branch and loads the history. Your working directory has to be clean, and you have to be signed in to the same claude.ai account. According to the docs, the handoff from the terminal only goes one way: you can't push a running terminal session to the cloud. Only the desktop app can do that, through its "Open in" menu.
A tip from the docs: plan locally and execute in the cloud. Start Claude in plan mode, save the plan as a file in the repository, push it, and then start claude --cloud with the instruction to carry out that plan.
Cloud environments: network, variables, setup scripts
Every cloud session runs in an environment, a saved configuration of network access, environment variables and setup script. There are four levels of network access:
Whatever the level, the session can reach GitHub through a separate proxy, the MCP connectors you enable, and the Anthropic API for Claude Code's own requests. According to the docs, data can therefore leave the VM even at "None".
Set your expectations accordingly: a cloud session starts from a fresh clone. What is in the repository is there: CLAUDE.md, .claude/settings.json with hooks and permission rules, .mcp.json, .claude/skills/, .claude/agents/. What only lives on your machine is missing: your personal ~/.claude/CLAUDE.md, your skills under ~/.claude/skills/, plugins, MCP servers from your user settings. A cloud session also does not install plugins that the repository turns on in .claude/settings.json. Skills you have enabled on claude.ai, however, load automatically.
Keys don't belong in environment variables, because according to the docs anyone who uses the environment can read them. For keys there are "network secrets": you store them on the environment, and a proxy attaches them to requests for the hosts you specify, without the key entering the VM. According to the docs, the VM has roughly 4 vCPUs, 16 GB of RAM and 30 GB of disk; after a few minutes without activity it pauses, and a paused VM can later be reclaimed entirely. Your conversation history is kept in that case, running background work is not.
Scheduled tasks, formerly routines
According to the docs, a scheduled task is a saved Claude Code configuration of a prompt, one or more repositories and connectors that runs automatically as a cloud session. Anthropic renamed it from "routines" to "scheduled tasks"; the old name still appears in a few places, such as the address claude.ai/code/routines and the /routines command. According to the docs, the feature is a research preview.
A task starts from one or more triggers:
- Schedule: hourly, daily, weekly or once at a specific time. Custom cron expressions work too, but not more often than once an hour, and a bare cron expression is in UTC.
- API: an HTTP POST to the task's own endpoint with a bearer token, for example from a deploy pipeline.
- GitHub: a repository event, such as a new pull request or a release.
You create them in the browser at claude.ai/code/routines, in the desktop app, or in the terminal with /schedule, for example /schedule daily PR review at 9am. A run in the cloud starts in auto mode while it is available and only stops when an action still needs approval.
Two things to know up front. First, a scheduled task belongs to your personal account: commits, pull requests and messages sent through connectors appear under your name. Second, each way of starting a run has an hourly limit, including 100 scheduled runs per hour per account.
Claude Code has three ways to schedule work. The docs compare them like this:
On my server, recurring jobs run as claude -p on schedules. For tasks without local files and credentials, the cloud version would be the alternative, with no machine that has to stay on.
What Claude Code costs in the cloud
There is no separate price list for the cloud. According to the docs, every cloud session counts toward your plan's usage limits, alongside the rest of your Claude and Claude Code usage, and there is no separate charge for the cloud VM. Parallel tasks use up your limits correspondingly faster.
The same applies to scheduled tasks. When a scheduled task hits your usage limit, further runs only go ahead if you have turned on usage credits, billed as metered overage; otherwise Claude Code rejects them until your window resets. On Team and Enterprise, an admin turns this on for the organization.
Either way, you need a Claude subscription with a claude.ai sign-in. With an Anthropic Console API key, or through Amazon Bedrock, Google Cloud or Microsoft Foundry, there are no cloud sessions and no cloud scheduled tasks. As an alternative there, the docs name GitHub Actions or GitLab CI/CD.
Limits you should know
- GitHub is required for cloning and pull requests. You can send a GitLab or Bitbucket repository as a bundle if you set
CCR_FORCE_BUNDLE=1, but you can't push results back to that remote.
- Organizations with an IP allowlist get an authentication error for sessions on Anthropic's infrastructure until Anthropic support sets up an exemption. Organizations with Zero Data Retention, the HIPAA configuration or customer-managed encryption keys (CMEK) can't use cloud sessions.
- Commands that only run in the terminal interface, such as
/plugin or /resume, aren't available in the cloud.
- Time limits: a foreground command runs for up to 2 minutes by default, up to 10 on request, then in the background for up to 30 more minutes. A setup script that takes longer than about five minutes isn't cached.
- No substitute for your local environment: interactive sign-ins such as AWS SSO don't work in the cloud.
For pull requests there is also "Auto-fix": Claude watches a pull request and responds to failing checks and review comments. That is not the same as the automated code review for teams that I describe in Claude Code Review: Cost, Setup, Limits.
Cloud session, Remote Control or parallel sessions on your machine?
- Cloud session, when the task is clear, lives in a GitHub repository and should keep running even when your computer is off.
- Remote Control, when you are in the middle of local work and just want to continue from another device.
- Several sessions on your machine, when the tasks need local files, tools or credentials. How to steer them from one screen is covered in Claude Code Agent View: Sessions Without tmux.
For larger efforts there are also projects: one ongoing conversation at claude.ai/code in which Claude starts and tracks parallel cloud sessions itself. According to the docs, projects are a public beta on Pro and Max, rolling out gradually, and are not yet offered on Team or Enterprise.
Frequently asked questions
What is Claude Code on the web?
The cloud version of Claude Code at claude.ai/code. The session runs on a VM that Anthropic runs by default, works on a cloned GitHub repository and keeps running when you close your browser.
How much does Claude Code in the cloud cost?
It has no price of its own. Cloud sessions and scheduled tasks count toward the usage limits of your Pro, Max, Team or Enterprise plan; according to Anthropic there is no charge for the VM. When a scheduled task hits your limit, it only keeps running with usage credits turned on.
What is the difference from Remote Control?
In a cloud session, Claude Code runs on Anthropic's infrastructure and does not need your computer. With Remote Control it runs on your computer, with your local files and tools; the browser and the app are just the window into it.
What happened to routines?
They are now called scheduled tasks. The feature is the same: a saved task that runs as a cloud session on a schedule, via API or on a GitHub event.
Can I build a website with Claude Code on the web?
Yes, if the website's code lives in a GitHub repository: Claude works on it like on any other project and pushes the changes back as a branch. Claude Code is not a builder that gives you a finished website without a repository. If that is what you are looking for, my comparison AI Website Generators Compared helps.
Do I need GitHub for Claude Code on the web?
For the normal route, yes: the session clones from GitHub and pushes back there. From the terminal you can also upload a local repository as a bundle with claude --cloud, for a GitLab or Bitbucket remote with CCR_FORCE_BUNDLE=1 set; the session then can't push results back to another remote.
Where to go from here
Start with a small, well-defined task in a repository where nothing important can break, and look closely at the diff in the browser before you merge a pull request. Leave the environment on "Trusted" for now and store keys only as network secrets. If you have not set up Claude Code locally yet, start with Install Claude Code: Windows, Mac, Linux. How others split up their Claude Code sessions is what we discuss in my German-speaking community Claude Practitioners.
