
Claude Code started life as a thing you ran in a terminal window on your own computer. You opened a folder, typed claude, and the agent worked on the files in front of it. When you shut the laptop, the session stopped. That description is still completely accurate — for local sessions.
But there is now a second kind of session, and it does not live on your machine at all. Anthropic calls it a cloud session, and the easiest way to misunderstand it is to assume it’s a web-based version of the app you already know.
It isn’t. It is the same agent, running somewhere else, reachable from five different places — and because it runs somewhere else, it has a different security shape, a different cost shape, and a habit of continuing to work after you’ve walked away from the computer. This guide explains what a cloud session actually is, the five doors into one, the three caveats Anthropic puts in its own documentation, and the handful of things that trip people up.
If you’re new to Claude Code entirely, start with What is Claude Code? — this article is the sequel to it, not a correction of it.
☁️ What a Cloud Session Actually Is
Anthropic’s definition is refreshingly flat. From the Claude Code cloud documentation, as it read on 18 October 2026:

“A cloud session is a Claude Code session that runs on cloud infrastructure instead of on your machine.”
That’s it. Same agent, same conversation, same tools — different computer. The agent is running on a virtual machine (a software-simulated computer rented by the minute) that Anthropic operates, or on your employer’s own hardware if your organisation has set that up.
The consequence is the bit worth internalising:
“The session keeps running after you close your laptop, and you can check on it or steer it from any device.”
A local Claude Code session is a process on your computer. If the lid goes down, the Wi-Fi drops, or the battery dies, the work stops. A cloud session is a thing on the internet that you have an account-level claim to. Starting one from your terminal and then checking on it from your phone on the train is the normal case, not a trick.
And it is not free in the sense people usually mean. The documentation is precise about this:
“It counts toward your plan’s usage limits alongside the rest of your Claude and Claude Code usage, and there’s no separate charge for the cloud VM.”
Those are two different claims and it’s worth keeping them apart. There is no extra line item for renting the machine. There is still a cost — it comes out of the same usage allowance as everything else you do with Claude.
🚪 The Five Doors
Here is where “web IDE” falls apart as a mental model. The browser is one way in. There are five, and the documentation lists them as equals:
| Surface | How you get in |
|---|---|
| Browser | claude.ai/code, also called Claude Code on the web |
| Mobile | The Code tab in the Claude app |
| Desktop app | Select Cloud instead of Local when you start a session |
| Terminal | claude --cloud |
| Routines | Scheduled and triggered runs each run as a cloud session |
Read that list again with the last row in mind. A routine — a job that fires on a schedule, or when something happens in your repository — is a cloud session too. Nobody is sitting in a browser tab watching it. The session is the unit; the interface is incidental.
That’s why “it’s a web IDE” is the wrong frame, and it’s the frame almost every first encounter installs. You land in a browser, see a conversation and a diff, and conclude you’ve been given an online code editor. What you’ve actually been given is a machine in a data centre with an agent on it, plus five windows onto it.
The practical test: close the tab. A web IDE stops being useful when you close the tab. A cloud session carries on.
📦 What the Cloud Machine Can Actually See
This is the question that catches everyone out, and the answer is not “your folder.”

When you run claude --cloud from a repository, the documentation is explicit about what happens:
“The cloud VM clones your current directory’s GitHub remote at your current branch, not your local checkout, so push first if you have local commits.”
So the cloud session reads your code from GitHub, not from your hard drive. If you’ve written something brilliant and haven’t pushed it, the cloud session has never heard of it. This is the single most common source of “why is Claude working on the old version?” confusion, and it follows directly from where the session lives.
There is a fallback. If your repository has no git remote at all, or it’s on GitHub but the Claude GitHub App isn’t installed on it, Claude Code bundles up your local repository and uploads it instead. That path has real limits, all of them from the documentation:
- the directory must be a git repository with at least one commit;
- the bundle must be under 100 MB (larger repositories degrade to the current branch only, then to a squashed snapshot, then fail);
- untracked files are not included — you have to
git addanything you want the session to see; - on macOS, Linux and WSL it needs git 2.31 or later;
- you can force this route with the environment variable
CCR_FORCE_BUNDLE=1.
That last one matters more than it looks. Repository cloning and pull-request creation require GitHub, so if your code is on GitLab or Bitbucket, the bundle is your only way in — and the session can’t push results back to that remote.
⚠️ The Three Caveats Anthropic Publishes Itself
The most useful criticism of a product is usually the criticism the vendor has already written down. There are three here, and they’re in the official docs rather than in anyone’s hot take.

1. “Network access disabled” does not mean data cannot leave
You can configure a cloud environment — the saved configuration that controls network access, environment variables and setup scripts — to restrict what the session can reach, up to and including turning network access off. It’s tempting to read that as a sealed box. The documentation says otherwise:
“When running with network access disabled, Claude Code can still communicate with the Anthropic API, which may allow data to exit the VM.”
Which makes sense the moment you think about it — an agent that cannot reach the model cannot think. But “network disabled” is being asked to carry a meaning it does not have, and that gap is exactly where people put code they shouldn’t.
2. The credential protection has a platform hole
When Claude Code bundles and uploads your local repository, it tries to keep your secrets out of the upload. On most platforms:
“Claude Code leaves uncommitted changes to files named like credentials or keys out of the upload. This includes
.envfiles, Terraform*.tfvarsfiles, and key files such asid_rsaand*.pem.”
That is a genuinely thoughtful safeguard. It does not apply everywhere. On native Windows:
“uncommitted changes to tracked files upload as they are, whatever the file’s name. Stash or revert an edit you don’t want in the cloud session before you start it.”
A protection that silently doesn’t apply on one platform is worse than no protection, because people learn the rule somewhere and carry it across. If you are on Windows and not running inside WSL, that filter is not running for you. This is the single most load-bearing fact in this whole series, and it gets a full article of its own below.
3. Running things in parallel costs proportionately more
Cloud sessions are so easy to start in parallel that the obvious move is to start four. The documentation declines to pretend this is free:
“cloud sessions share rate limits with all other Claude and Claude Code usage within your account. Running multiple tasks in parallel consumes more rate limits proportionately.”
“Proportionately” is doing honest work in that sentence. Four parallel sessions is four sessions’ worth of usage, drawn from the same allowance that your browser chats and your local terminal sessions draw from.
🔀 Moving Work Between Your Terminal and the Cloud
There are three similar-looking flags here and they do genuinely different things. The documentation disambiguates them explicitly, which is usually a sign that support has been fielding the question.

--cloud starts a new cloud session. The older --remote spelling still works as a deprecated alias for it.
--remote-control is a different feature entirely:
“
--cloudcreates cloud sessions.--remote-controlis unrelated: it lets you monitor and steer a local CLI session from claude.ai or the Claude app.”
So Remote Control leaves the work on your machine and gives you a window onto it from your phone. --cloud moves the work off your machine entirely. They solve adjacent problems and are not interchangeable.
--teleport pulls an existing cloud session down into your terminal — and this is the one people get wrong, because the obvious reading is that it moves the session. It doesn’t. It forks it:
“The terminal gets its own copy of the session: new work there stays local and doesn’t appear in the cloud session on claude.ai or the Claude mobile app.”
After a teleport you have two copies of the conversation, and they diverge from that moment. Anything you do locally stays local. If you want to keep steering from your phone after teleporting, you start Remote Control in the local session — the documentation says so directly.
And for completeness, because the names invite it: “--teleport is distinct from --resume.” --resume reopens a conversation from your machine’s own local history and doesn’t list cloud sessions at all.
One structural asymmetry worth knowing before you plan a workflow around it:
“From the CLI, session handoff is one-way: you can pull cloud sessions into your terminal with
--teleport, but you can’t push an existing terminal session to the cloud.”
You can start a new cloud session from the terminal, and the Desktop app can send a local session to the cloud from its Open in menu. But there’s no terminal command that takes the conversation you’re in the middle of and relocates it upward.
You can, however, talk to a running cloud session from anywhere you’re logged in, with claude -p "your message" --cloud <session-id>. The CLI sends no local session state, so the command doesn’t need to run on the machine that started the session.
🔒 Who Can’t Use It
Cloud sessions are available on Pro, Max and Team plans, and for Enterprise users with premium seats or Chat + Claude Code seats. The exclusions are more interesting than the inclusions, because they’re all organisational and none of them are obvious from the pricing page.

Organisations with Zero Data Retention enabled, or with the HIPAA configuration applied, can’t use the /web-setup connection command or other cloud session features. Organisations using customer-managed encryption keys (CMEK) “can’t start cloud sessions either.”
And one that produces a genuinely baffling error message if you don’t know about it:
“If your organization has IP allowlisting enabled, every Anthropic-hosted cloud session fails with an authentication error.”
The reason is structural rather than a bug: Anthropic-hosted cloud sessions call the API from Anthropic’s infrastructure, not from your office network, so your allowlist doesn’t recognise them. If your employer locks Claude to a range of IP addresses and cloud sessions mysteriously won’t authenticate, that’s the cause — and it reads as a login problem, which sends people down entirely the wrong path.
The escape hatch is a self-hosted environment, where sessions run on your own infrastructure and call the API from your own network. That’s a real option, with a real trade: in a self-hosted environment, isolation is your deployment’s responsibility, your deployment supplies the git credentials, and you restrict session egress at your own network boundary. The convenience of the managed version is precisely the part you give up.
👥 Sharing, and What a Shared Session Contains
Sharing behaviour depends on your account type, and the difference between them is worth a second look.
On Team and Enterprise accounts, visibility is Private or Team, and repository-access verification is enabled by default — recipients are checked against the GitHub account connected to their Claude account.
On Max and Pro accounts, visibility is Private or Public, where public “makes the session visible to any user logged into claude.ai” — and the verification posture is reversed. The documentation states plainly that for Max and Pro, “Repository access: verification isn’t enabled by default,” followed by this:
“Sensitive content: check your session before sharing. Sessions may contain code and credentials from private GitHub repositories”
Sharing is a toggle you operate, not something that happens to you — sessions are not public by default. The thing to be careful about is what’s inside a session you’ve decided to share, and the fact that on a personal plan nobody is checking whether the recipient had access to that repository in the first place. Both of those settings can be changed under Settings → Claude Code → Sharing settings.
⏱️ Sessions Expire, and What Survives
Cloud sessions stop after a period of inactivity and the VM is reclaimed. Reopening the session from claude.ai/code provisions a fresh machine, and the split is clean:
- Restored: your conversation history.
- Not restored: background work that was still running when the VM was reclaimed — subagents, shell commands, and the pending wakeup of a self-paced loop.
There’s a subtle trap in the inactivity clock: a session counts as inactive while it waits for you to approve a tool call or sign in to a connector. So a session can expire sitting on a question it asked you. If you walk away from a prompt, you may come back to a session that needs reprovisioning rather than one holding its place.
🗺️ The Rest of This Series
This article is the map. Five companion pieces go deeper on the parts that deserve their own treatment, and they publish over the coming weeks:

What your cloud session can and can’t reach takes the three caveats above and works through them properly — especially the Windows credential-filter asymmetry, which is the one fact here most likely to cost somebody something real. If you read only one of the five, read that one.
Plan locally, execute in the cloud covers the workflow the documentation itself recommends, rather than one we invented: use plan mode to collaborate on the approach while Claude reads files and explores without editing anything, save the plan into the repository, commit and push it, then start a cloud session to execute it. It’s a good answer to “how much autonomy should I hand over?” — you supervise the thinking and delegate the typing.
Running work in parallel is about the thing cloud sessions make almost too easy, and the honest arithmetic underneath it. Several independent sessions at once is a genuinely new capability; it also draws on one shared allowance, proportionately.
Cloud vs local vs Remote Control untangles the three things sketched above. The documentation disambiguates them in two separate places, which tells you how often they get confused — and the teleport-is-a-fork behaviour catches careful people, not careless ones.
Auto-fixing pull requests covers the loop where Claude watches a pull request, reacts to failing CI and reviewer comments, and pushes fixes. It includes the detail that most deserves to be read before you switch it on: the replies are posted using your GitHub account. The documentation puts it like this —
“Claude may reply to review comment threads on GitHub as part of resolving them. These replies are posted using your GitHub account, so they appear under your username, but each reply is labeled as coming from Claude Code so reviewers know it was written by the agent and not by you directly.”
— along with a specific warning that if your repository uses comment-triggered automation such as Atlantis, Terraform Cloud, or GitHub Actions that run on issue_comment events, Claude replying on your behalf can trigger those workflows. Auto-fix requires the Claude GitHub App installed on the repository.
Two features sit outside this series but are worth knowing exist: Projects, a single conversation where Claude coordinates several cloud sessions across your repositories and reports back, and the web quickstart, which walks a first session end to end.
❓ Frequently Asked Questions
Is a cloud session the same as Claude Code on the web? Claude Code on the web is one surface for cloud sessions — the browser at claude.ai/code. The session is the underlying thing; the browser, the mobile Code tab, the Desktop app’s Cloud mode, claude --cloud, and routines are five ways to reach it.
Does a cloud session cost extra? There’s no separate charge for the cloud VM, but the session consumes your plan’s usage limits alongside everything else you do with Claude. Running several in parallel uses proportionately more.
Can a cloud session see files I haven’t pushed to GitHub? Generally no. The cloud machine clones your GitHub remote at your current branch, not your local checkout — so push first. The exception is the bundle-upload path, which applies when there’s no git remote or the Claude GitHub App isn’t installed on the repository.
If I teleport a session to my terminal, does the cloud copy stop? No, and this is the common misunderstanding. Teleport gives the terminal its own copy. New work in the terminal stays local and doesn’t appear in the cloud session on claude.ai or the mobile app.
I’m on Windows. Is the secret-file protection running for me? Not on native Windows. The filter that excludes uncommitted changes to .env files, *.tfvars, id_rsa and *.pem from a bundle upload applies on macOS, Linux and WSL. On native Windows, uncommitted changes to tracked files upload as they are, regardless of filename — stash or revert anything you don’t want sent.
Why do my cloud sessions fail to authenticate at work? If your organisation uses IP allowlisting, every Anthropic-hosted cloud session fails with an authentication error, because those sessions call the API from Anthropic’s infrastructure rather than your network. Zero Data Retention, the HIPAA configuration, and customer-managed encryption keys also block cloud sessions.
Does Claude Code still only work in the folder I open it in? For local sessions, yes — that’s still the model. A cloud session is a different arrangement: the agent works on a clone of your repository on a machine in a data centre, inside a configured cloud environment that governs what it can reach.
Everything quoted above was read directly from Anthropic’s Claude Code cloud documentation at code.claude.com/docs/en/claude-code-on-the-web on 18 October 2026. Version-gated details and plan availability move quickly on a feature this new — check the page itself before relying on a specific number.
Start here if you’re new: What is Claude Code?
📦 Recommended Resources
Some links below are Amazon affiliate links. If you buy through them we earn a small commission at no extra cost to you.
Start with the free primary material, because on this subject the vendor’s own documentation is both the most accurate source and the most candid one. Use Claude Code in the cloud is the page everything in this article is built on, and it is worth reading in full rather than in search results — the caveats that matter are not in the summary. Pair it with Cloud environments, which is where network access, environment variables and setup scripts are actually configured, and Remote Control, which exists to stop people confusing two different features. All three are free, and they change often enough that a book will never be the right place to learn the current behaviour.
The three books below are for the layer underneath — the reasoning that makes a feature like this make sense, rather than the feature itself.
- Software Engineering at Google by Titus Winters, Tom Manshreck and Hyrum Wright — the best book on what changes when engineering tooling runs somewhere other than a developer’s laptop. Its central distinction, between programming and engineering sustained over time, is exactly the lens that makes a cloud session interesting: not “I can code in a browser” but “the work outlives the machine and the session, and therefore needs different guarantees.” The chapters on build systems and automation are the relevant ones here.
- Accelerate: The Science of Lean Software and DevOps by Nicole Forsgren, Jez Humble and Gene Kim — the evidence base for why shortening the loop between intent and running code matters more than most teams assume. Read it before you decide whether moving work off your machine is worth the change in habits; it is the book that distinguishes “feels faster” from “is measurably faster,” which is the question this whole feature poses.
- Pro Git by Scott Chacon and Ben Straub — unglamorous and completely load-bearing. A cloud session clones a remote, works on a branch, and pushes it back; every confusion people have with the feature turns out to be a confusion about remotes, branches and what has actually been pushed. Also free to read in full at git-scm.com/book — buy it only if you prefer paper.
✅ Pro tip: Before you start your first cloud session, run git status and git log origin/<your-branch>..HEAD in the project you intend to use. The first tells you what is uncommitted; the second tells you what is committed but not pushed. The cloud VM sees neither. Two commands, and they prevent the single most common frustration with this feature.