Claude Code Cloud (2026): How It Works, Limits, Options

Do not index

Search for "Claude Code cloud" and nearly everything points at one product: Claude Code on the web, Anthropic's research preview that runs coding sessions on its own managed VMs at claude.ai/code. It is included with Pro, Max, and Team plans, it shares your plan's usage limits, and there is no separate compute charge for the VM. It is also deliberately task-scoped: idle sessions expire, the VM is reclaimed, and anything still running inside is gone. If what you actually want is Claude Code on a cloud computer that stays on around the clock, that is a different thing called a persistent instance, and it is what we run at Agent37 (always-on, from $1.99/mo shared or $4.94/mo dedicated).

This guide covers both lanes with September 2026 details: how Anthropic's cloud sessions work, which plans include them, the scheduled "routines" feature, the limits people actually hit, and when a persistent instance is the better tool.

What "Claude Code cloud" actually means

There is no Anthropic product literally named "Claude Code Cloud". When people say it, they mean one of three things:

Claude Code on the web. Anthropic's official cloud product: you delegate a coding task in the browser at claude.ai/code, and it runs on an Anthropic-managed VM. The CLI flag is literally claude --cloud (the older --remote spelling still works as a deprecated alias), and the docs call them cloud sessions.

Routines. Scheduled cloud runs of Claude Code: cron-style triggers, an HTTP endpoint per routine, or GitHub events. Same cloud infrastructure, no laptop involved.

Running Claude Code on a cloud server you control. A VPS, a Docker host, or a managed always-on instance where Claude Code is installed like it would be on your laptop, except the machine never sleeps.

The first two are Anthropic-hosted and task-scoped by design. The third is your infrastructure and your rules. Most confusion around this keyword comes from mixing them up, so let's take them in order.

Claude Code on the web launched as a research preview on October 20, 2025 for Pro and Max, and was extended to Team and Enterprise premium seats a few weeks later. As of September 2026 it is still labeled a research preview.

How a Claude Code cloud session works

Every cloud session gets a fresh, isolated VM running Ubuntu 24.04 with roughly 4 vCPUs, 16 GB of RAM, and 30 GB of disk. The environment comes preloaded with the usual toolchain: Python, Node, Ruby, PHP, Java, Go, Rust, C/C++, Docker with compose, PostgreSQL 16, Redis 7, git, and the GitHub CLI.

A few design choices define the experience:

GitHub runs through a credential proxy. Your real GitHub credentials never enter the VM. The session clones through Anthropic's proxy, git push works only against the session's own working branch, and pull request creation is built in.

Network access is allowlisted. Each cloud environment has one of four access levels: None, Trusted (a default allowlist of package registries, GitHub, and cloud SDK endpoints), Full, or Custom. All outbound traffic passes through a security proxy.

You never get a shell. There is no SSH into the VM and no terminal tab. Claude runs every command for you, and you review what it did.

Setup is cached, processes are not. A setup script can install packages or pull Docker images, and the resulting filesystem snapshot is reused for about 7 days. Running processes never survive between sessions.

Secrets can stay outside the VM. On Pro and Max plans, per-environment API credentials are attached to matching outbound requests by Anthropic's proxy, so Claude never sees the raw keys. Team and Enterprise plans do not have API credentials yet, so keys there live in environment variables that anyone using the environment can read.

Sessions are device-independent: close your laptop and the session keeps running, and claude --teleport can pull one down into a local terminal later. How the browser, SSH, and vendor-dashboard access paths compare is its own topic, and we cover it in running Claude Code remotely.

With the Claude GitHub App installed, cloud sessions can also auto-fix pull requests: a session subscribes to PR webhooks and pushes fixes for CI failures and review comments on its own.

Which plans include Claude Code cloud, and what it costs

Claude Code on the web is included with the paid claude.ai plans: Pro at $20/month ($17/month billed annually), Max 5x at $100/month, Max 20x at $200/month, and Team, plus Enterprise premium seats and Chat + Claude Code seats. It is not on the free plan, and it does not work with a Console API key or through Bedrock or other third-party providers. Organizations with Zero Data Retention enabled are excluded entirely.

The pricing model is simple: there is no separate compute charge for the cloud VM. Cloud sessions draw from the same usage pool as everything else on your account, the familiar 5-hour session window plus a weekly limit. Run three cloud sessions in parallel and you burn limits roughly three times as fast. When you hit a limit, you can either wait for the reset or turn on usage credits, which bill overflow at standard API rates.

That "no compute charge" framing is honest but worth reading carefully: you are paying with your plan's capacity. A Pro plan that comfortably covers an hour or two of terminal Claude Code per day can feel tight once cloud sessions and routines start drawing from the same pool. The full cost picture, including the API-billed path, is in what Claude Code actually costs.

Defaults differ by plan, too: Max, Team premium, and API accounts default to Opus 5, while Pro and Team standard default to Sonnet 5.

Routines: scheduled Claude Code in the cloud

Since April 2026, Claude Code on the web includes routines, a research-preview scheduler. A routine runs Claude Code in the cloud on a recurring cadence (hourly, nightly, weekly), on demand via a per-routine HTTP endpoint, or in response to GitHub events like a PR opening. You manage them at claude.ai/code/routines or with /schedule in the CLI.

Three constraints matter in practice:

One hour is the floor. Sub-hourly cron expressions are rejected.

There is a daily cap on routine runs per account, and runs are rejected once you hit plan limits unless usage credits are on.

Every run is a fresh session. The repo is cloned from the default branch each time; nothing carries over from the previous run except what you committed.

Routines are excellent for "check CI every morning and fix what broke" work. They are not a substitute for a process that needs to stay alive, hold state in memory, or react in seconds.

The limits people actually hit

None of these are bugs. They are the shape of a task-scoped product:

Sessions expire. After a period of inactivity the VM is reclaimed. Reopening restores your conversation history on a fresh VM, but subagents, shell commands, and anything else still running when the VM was reclaimed is not restored. There is no way to keep a daemon alive.

GitHub or nothing for output. Cloning and PR creation require GitHub. GitLab and Bitbucket repos can be sent up as a local bundle, but the session cannot push results back to them.

Chat channels are not hosted. The Telegram, Discord, and iMessage channels in Claude Code's docs run on your own machine: they push messages into a session your computer keeps open. Slack is the one hosted exception, where mentioning @Claude spawns a cloud session.

Shared limits cut both ways. A runaway cloud session eats the same weekly budget as your chat usage and your terminal sessions.

No shell. You cannot SSH in to poke at state. Claude can install packages mid-session, but those installs do not persist beyond the session, so anything permanent goes through the setup script.

Corporate networks can break it. Organization-level IP allowlists break Anthropic-hosted sessions by default; Anthropic support can exempt its hosted services from the allowlist, and Team and Enterprise can use self-hosted environments instead (a public beta since August 2026).

If your use of Claude Code fits inside "delegate a task, get a PR", these limits will rarely bother you. The moment your workflow needs a machine, they define it.

Cloud sessions vs. an always-on Claude Code instance

Here is the honest comparison, side by side:

ㅤ	Claude Code on the web	Persistent instance (Agent37)
Compute cost	$0 extra on top of your plan	From $1.99/mo shared, $4.94/mo dedicated
Session lifetime	Task-scoped, idle VMs reclaimed	Runs until you delete it
Scheduling	Routines, 1-hour minimum	Any cadence, cron included
Git hosts	GitHub only for push and PRs	Any remote you configure
Shell access	None, Claude runs commands	Full web terminal, plus a file browser
Long-running processes	Not restored after reclaim	Stay up 24/7
State between tasks	Repo snapshot and history only	Whole filesystem persists
Plan usage	Draws from your Claude plan limits	Your Anthropic account or API key, your choice

Use Claude Code on the web first for what it is built for. If your work is PR-shaped, lives on GitHub, and fits in discrete tasks, it is effectively bundled with the plan you already pay for, and it is very good.

Reach for a persistent instance when you cross one of these lines:

You need something to keep running: a dev server under test, a long migration, a watcher, a bot.

You want schedules tighter than an hour, or jobs that build on the last run's state.

Your repos live on GitLab, Bitbucket, or a private git host.

You want to open a real terminal on the box, install what you like, and keep it warm.

You are provisioning Claude Code per client or per teammate and need each one isolated on its own machine.

For a full ranking of the server options themselves, DIY VPS included, see the best Claude Code server setups.

Running Claude Code on a computer that stays on

This is the lane Agent37 builds for. Claude Code hosting on Agent37 Cloud gives you a persistent Linux instance with Claude Code preinstalled, provisioned from one API call:

curl -X POST https://api.agent37.com/v1/instances \
  -H "Authorization: Bearer $AGENT37_KEY" \
  -H "Content-Type: application/json" \
  -d '{"template": "agent37-claude-code"}'

The instance comes up at its own URL with Claude Code as the default agent, a full TTY web terminal, and a file browser. It is a lean image (no browser or desktop), and it runs on your own Anthropic account: sign in with claude auth login in the instance terminal, drop in a long-lived token from claude setup-token, or set an ANTHROPIC_API_KEY. Model usage bills your Anthropic account directly, not your Agent37 balance. You can even point it at a different provider entirely via ANTHROPIC_BASE_URL, OpenRouter included.

Because the machine persists, the things the web product scopes out just work: cron schedules at any cadence, processes that stay up, any git remote, app integrations (Gmail, Slack, Notion and hundreds more via Composio's MCP server, registered with Claude Code automatically), and files that are still there next week.

Pricing is boring on purpose: the 2 vCPU / 4 GB shape is $4.94/month dedicated, or from $1.99/month on shared capacity, metered per minute from a prepaid balance with no seats. New workspaces get a $1 starter credit (about a week of the cheapest instance, no card required). Agent37 is backed by Y Combinator and SOC 2 Type I audited, with Type II in progress. One honest caveat: at these prices there is no human support hotline; an AI bot filters support first.

Prefer to build it yourself? Our open-source claude-code-docker image gives you the same pinned, non-root container with persistent OAuth login to run on any host, and the always-on options for the Agent37 Cloud API are documented if you outgrow one box.

Claude Code cloud FAQ

Is Claude Code on the web free?

No. It requires a paid claude.ai plan: Pro ($20/month, $17 annually), Max, Team, or an Enterprise premium or Chat + Claude Code seat. Within those plans, cloud sessions add no separate compute charge; they draw from your plan's usage limits.

Can Claude Code run 24/7 in Anthropic's cloud?

No. Cloud sessions are task-scoped: idle VMs are reclaimed and background processes are not restored. Routines get you recurring runs down to hourly, but between runs nothing is alive. For a Claude Code box that is actually on 24/7, you need a persistent instance you control.

Does Claude Code cloud work with GitLab or Bitbucket?

Partially. You can send a non-GitHub repo into a session as a local bundle, but the session cannot push results back. Cloning and pull request creation require GitHub. On a persistent instance, any git remote works.

Can I move a cloud session to my terminal?

Yes. claude --teleport gives you an interactive picker and pulls the session down into your local terminal. From the CLI the handoff is one-way: you cannot push an existing terminal session up to the cloud, though the Desktop app's Continue in menu can send a local session to the web.

What does a persistent Claude Code instance cost?

On Agent37, the 2 vCPU / 4 GB shape is $4.94/month dedicated or from $1.99/month shared, plus model usage on your own Anthropic account or API key. A DIY VPS runs $4 to $12/month before your time; the server setups guide compares the options.

The bottom line on Claude Code cloud

"Claude Code cloud" is really two products wearing one name. Claude Code on the web is Anthropic's task-runner: managed VMs, GitHub in and out, no extra compute charge, sessions that end when the work does. It is the right default for delegated, PR-shaped tasks, and it is already in your plan.

A persistent instance is the other product: a cloud computer that is yours, where Claude Code stays installed, logged in, and running, on any schedule, against any git host, with state that survives the week. That is what Agent37's Claude Code hosting does: always-on from $1.99/month shared or $4.94/month dedicated, deployed in one API call, with a $1 credit to try it. When your Claude Code workflow stops fitting inside a task, give it a computer that stays on.

Claude Code Cloud, Explained (2026): How It Works, Its Limits, and the Always-On Alternative