twimco.Let’s talk ↗︎
Menu
Field notes / Reviews

Claude Code cloud environments review: worth it when secrets stay outside the VM

Claude Code cloud sessions run on isolated VMs with saved environments for network access, setup scripts, and network secrets that stay outside the sandbox. The buying question is whether your team will configure that boundary—or paste keys into the prompt.

Editorial engraving of a sunny engineering loft where a staff engineer types beside a sealed envelope while a teammate notes a paper allowlist.

Claude Code cloud sessions (Pro, Max, Ultra, Team, and eligible Enterprise seats) run on Anthropic-managed or self-hosted VMs you can start from the browser, mobile, Desktop, or `claude --cloud`. Saved cloud environments control network access, env vars, and setup scripts; network secrets are attached by a proxy after requests leave the VM so the key never sits in the sandbox. Worth enabling for parallel repo tasks and auto-fix when GitHub access and secret handling are deliberate; weak if the team treats the cloud VM like a laptop with .env files.

What Claude Code cloud sessions are

Claude Code can run a session on cloud infrastructure instead of on your laptop. The session keeps going after you close the machine. You start one from claude.ai/code in the browser, from the Claude mobile app's Code tab, from the Desktop app with a Cloud environment selected, from the terminal with claude --cloud, or as a scheduled cloud task.

Cloud sessions are available on Pro, Max, Ultra, and Team plans, and for Enterprise users with premium seats or Chat + Claude Code seats. They share your plan's usage limits; Anthropic does not describe a separate charge for the cloud VM itself. Each Anthropic-hosted session runs in an isolated VM. Organizations can also route sessions to self-hosted environments.

Every session runs inside a cloud environment: the saved configuration for network access, environment variables, and setup scripts. Network secrets are different from ordinary env vars. In Anthropic-hosted environments, keys you add as network secrets stay outside the sandbox; a proxy attaches them to matching outbound requests after the request leaves the VM, so the key is not sitting in the session filesystem for the model to read.

A cloud coding agent is only as trustworthy as the keys you refuse to put inside it.

Who it is for

Engineering teams that already use Claude Code and want parallel work on GitHub repositories without babysitting a local checkout. Product and platform groups that need a PR opened while someone is on a phone. Organizations willing to decide network allowlists and secret placement up front.

It is a weaker fit for Free-tier curiosity, for teams that cannot connect GitHub, and for regulated setups that block cloud sessions entirely—Zero Data Retention, certain HIPAA configurations, and customer-managed encryption keys are called out as incompatible with cloud sessions in the docs. Organization IP allowlists that do not exempt Anthropic-hosted infrastructure will also fail hosted sessions.

Decision guide

Configure the boundary vs hope the prompt behaves

ConsiderationConfigured cloud environmentLaptop habits in a VM
Secrets outside the VMNetwork secrets attached outside the sandboxAPI keys pasted into chat or committed .env
Network allowlistTrusted or custom network allowlistFull egress 'so installs work' with no owner
Cached setup scriptSetup script that finishes and cachesManual apt rituals every session
GitHub accessGitHub App or /web-setup with scoped accessPersonal tokens shared in a team channel
Decision framework: assess these conditions against your own engineering workflow.
A cloud coding agent is only as trustworthy as the keys you refuse to put inside it.

Where it is strong

Isolation by default. Separate VMs, network controls, and credential proxies are the right shape for unattended coding work.

Network secrets as a product idea. Keeping keys outside the sandbox is the feature to evaluate, not the demo that writes a function. If your threat model is "the agent saw the key," this architecture is the reason to prefer cloud environments over a laptop session with .env loaded.

Move work between surfaces. --cloud to start, --teleport to pull a finished session into a local terminal, mobile steering while a long task runs—those handoffs match how people actually work.

Auto-fix with eyes open. Claude can watch a PR for CI failures and review comments and push fixes when clear. That is powerful on application repos and dangerous on repos where a comment triggers infrastructure deploy automation. The docs warn about comment-triggered tools such as Atlantis-style flows.

Where it falls short—or needs your discipline

Setup still fails closed if you skip it. Dependencies, network, and keys only stay ready when someone configures the environment. Auto-detect announcements do not replace an allowlist owner.

Bundled local uploads have sharp edges. When Claude Code uploads a local repo instead of cloning GitHub, uncommitted sensitive files may or may not be excluded depending on platform. Native Windows behavior differs. Read that section before you --cloud from a dirty tree.

Session expiry drops in-flight work. Idle sessions reclaim the VM; conversation returns, but shell commands and subagents may need a rerun.

Not a substitute for review. Cloud sessions open PRs for humans. If your team merges without reading diffs, the cloud VM only accelerates the mistake.

How platform teams should evaluate it. Ask three questions in the security review. First: where do production secrets live, and can we prove they never enter the VM filesystem? Network secrets are the affirmative answer Anthropic documents; pasted keys in chat are the failure. Second: who owns the network allowlist when a package install needs a new host? Third: which repositories may enable auto-fix, given comment-triggered deploy tooling?

For Team and Enterprise, Owners control Quick setup and cloud-session permissions. Turn Quick setup on only when you intend web-setup and fewer prompts—not as a casual default. Self-hosted environments change isolation responsibility: your network boundary becomes the control, and network secrets as described for Anthropic-hosted VMs do not apply the same way.

A fair pilot is one repository, Trusted network, no production secrets, one cloud task that opens a PR, and a human merge. Expand to auto-fix only after CI noise is understood. Expand to network secrets only for hosts on a written list. Skip public-session sharing on Max, Ultra, or Pro until someone reads the session for credentials.

Claude Code cloud is a strong fit for teams that already review PRs seriously. It is a poor fit for teams that treat agent output as already approved.

Cost and attention, not just safety. Cloud sessions share plan usage limits with the rest of Claude and Claude Code. Parallel --cloud tasks multiply spend and review load. Set a team norm for how many concurrent cloud sessions a person may start, and require a human reviewer on every PR the cloud opens—especially when auto-fix is on.

Teleporting a session into a local terminal is powerful and easy to misuse on a dirty tree. Teach the clean-git requirement the same way you teach "do not commit secrets." The docs are explicit; your onboarding should be too.

For regulated buyers evaluating AI applications and agents, ask vendors—including yourselves—whether coding agents run with network secrets outside the compute, or with env files inside it. That question separates theater from design.

Keep a short internal allowlist of repositories approved for cloud sessions. Not every repo with a GitHub App install should get auto-fix or network secrets. Start with application code; delay infrastructure-as-code repos until comment-triggered automation is reviewed.

Plain verdict

Enable Claude Code cloud environments when you will put network secrets outside the VM, keep Trusted or custom network rules owned, and start with read-mostly tasks before auto-fix. Skip it—or keep it off org-wide—if the plan is to paste production keys into chat so "the agent can finish."

For help designing AI applications and agents with real containment, contact TWIMCO.

Related work: Spin:Market

Sequence of work

A sensible first week with cloud sessions

  1. Day 1

    Connect GitHub, create Default environment with Trusted network, add no production secrets yet.

  2. Day 2

    Run one read-mostly task with --cloud; review the branch and diff before any auto-fix.

  3. Day 3

    Add network secrets for the hosts the job truly needs; confirm keys never appear in session files.

  4. Day 5

    Enable auto-fix only on repos where a PR comment cannot deploy infrastructure.

Illustrative workflow, not a client screenshot or measured result.
twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎