skip to content

A Claude Code cloud session can't see your user-level MCP servers, skills or AWS SSO login — what does it actually inherit and reach?

level: seniorimportance: should knowfreq 11%

answer

  1. a different machine entirely
  2. committed files travel, home doesn't
  3. project scope for MCP
  4. Trusted network level by default
  5. routes that skip the allowlist

basics

~20 s

What the repository carries: CLAUDE.md, .claude/ skills, agents and rules, and in a one-repository session its settings.json and .mcp.json. User-level files, locally added MCP servers, plugins and interactive SSO never arrive, and outbound network defaults to the Trusted allowlist.

solid answer

~50 s

A cloud session starts from a **fresh clone** on a VM, so anything that lives only on your machine is missing: `~/.claude/CLAUDE.md`, `~/.claude/skills/` and `agents/`, servers added with `claude mcp add` at local or user scope, user-enabled plugins, device-delivered managed settings and browser-based logins such as AWS SSO. What is committed travels: `CLAUDE.md`, `.claude/rules/`, `.claude/skills/`, `.claude/agents/`, and in a single-repository session the hooks and permission rules in `.claude/settings.json` plus `.mcp.json`. Plugins a repository enables are not installed. Your organization's server-managed settings do arrive. Outbound traffic follows the environment's network level, **Trusted** in the starting `Default` environment, while GitHub goes through a proxy that keeps your credentials out of the VM. To make a workflow cloud-ready, commit it: `claude mcp add --scope project`, skills under `.claude/`, and secrets as environment variables or, on Pro and Max, API credentials on the environment.

code

bash · 7 lines
bash
# Re-add the tracker server at project scope so cloud sessions load it
claude mcp add --transport http tracker --scope project https://mcp.example.com/mcp

# .mcp.json is now in the repository root; commit and push it
git add .mcp.json
git commit -m "Share tracker MCP server with cloud sessions"
git push

go deeper

for a junior

Remember that a cloud session is a fresh clone on another machine, so only committed configuration such as CLAUDE.md and .claude/ files is there.

for a middle

Explain the carry-over rules: project files in a one-repository session, nothing from your home directory, no repository-enabled plugins, and the Trusted network level by default.

for a senior

Diagnose a cloud run that behaves differently from local by walking the layers: clone contents, settings scopes, server-managed policy, network level, and how credentials are attached without entering the VM.

for a principal

Treat the cloud environment as shared infrastructure: who can read its variables, which network level each team gets, and when a self-hosted environment is worth operating instead.

## A different machine, not your machine A **cloud session** is a Claude Code session that runs in a **cloud environment**: an isolated VM, Anthropic-managed by default, that starts from a **fresh clone** of your repository. That one fact explains almost every surprise. Anything you committed is there; anything you installed or configured only on your laptop is not. A **cloud environment** is the saved configuration that sets the VM's network access, environment variables and setup script; the one onboarding creates is called `Default`. ## What carries over | Item | In a cloud session? | Why | | :- | :- | :- | | Repository `CLAUDE.md`, `.claude/rules/` | Yes | Part of the clone | | Repository `.claude/skills/`, `.claude/agents/`, `.claude/commands/` | Yes | Part of the clone | | Repository `.claude/settings.json` hooks and permission rules | Yes, with one repository | Read from the clone the session starts in | | Repository `.mcp.json` | Yes, with one repository | Found from the working directory | | Plugins the repository enables | No | Cloud sessions do not install them | | `~/.claude/CLAUDE.md`, user skills and agents | No | They live on your machine | | MCP servers added at local or user scope | No | Written to `~/.claude.json` on your machine | | Interactive logins such as AWS SSO | No | They need a browser login the VM cannot do | A session with **several repositories** starts above the clones, so it does not read their permission rules, hooks or `env`; only a one-repository session loads `.mcp.json`. ## Settings and policy layers - **Project settings** (`.claude/settings.json`) apply in a one-repository session because the file is in the clone. - **User and local settings** (`~/.claude/settings.json`, `.claude/settings.local.json`) are never read. - **Managed settings delivered to your device** (a `managed-settings.json` file or MDM profile) do not reach the VM; **server-managed settings** from your organization's admin console do. - **`/config`** typed at claude.ai/code opens the Claude Code page of your account settings instead of changing a value; change a cloud setting through an environment variable on the environment or by committing it to the repository. - **Authentication** is always your subscription: an `ANTHROPIC_API_KEY` set in the cloud environment does not override it. - **Plugins** named in a committed `enabledPlugins` or `extraKnownMarketplaces` are not installed in the VM; skills you enable on claude.ai, by contrast, load automatically. ## Network and credentials Each environment has one **network access level**: | Level | Outbound connections | | :- | :- | | None | Nothing through the session's network | | Trusted (the `Default` environment's level) | Allowlisted domains: package registries, GitHub, cloud SDKs | | Full | Any domain | | Custom | Your own allowlist, optionally with the defaults | Some traffic does not pass through that allowlist at any level: 1. **GitHub**, through a separate proxy that attaches your credential server-side, so the token never enters the VM. 2. **MCP connectors** you enable, whose traffic goes through Anthropic's servers. 3. Hosts listed on the environment's **API credentials** (Pro and Max), which a proxy attaches to matching requests. 4. The **Anthropic API** itself, even at **None**, which is why the docs warn that data can still leave the VM that way. Environment variables and setup scripts are readable by anyone who uses the environment, so they are a poor home for secrets; on Pro and Max, store a key as an API credential instead. Anthropic-hosted VMs also have approximate ceilings (about 4 vCPUs, 16 GB of RAM and 30 GB of disk, subject to change). For heavier builds, the docs point you to Remote Control on your own hardware or to a self-hosted environment. ## Making a workflow cloud-ready 1. Move personal skills and agents you need into the repository's `.claude/` directory and commit them. 2. Re-add a local MCP server with `--scope project`, which writes `.mcp.json`, and commit that file. 3. Put build dependencies in the environment's **setup script** and the variables they need on the environment. 4. Pick the narrowest network level that works, and add domains with **Custom** rather than jumping to **Full**. 5. From the CLI, choose which environment `claude --cloud` uses with `/remote-env`. The underlying rule is simple: if a teammate cloning the repository on a clean machine would not get it, neither will the cloud session.

  • The cloud session needs an internal package registry that is not on the default allowlist. What do you change?
    Edit the cloud environment's **Network access** and choose **Custom**, adding the registry's domain and optionally keeping the defaults. Existing sessions in an Anthropic-hosted environment pick up the change within about a minute. Choosing **Full** also works but opens every domain, which is wider than the task needs.
  • Why is putting a production API key in the environment's variables a poor idea?
    Anyone who uses the environment can read its environment variables and setup script, and the VM process sees the value. On Pro and Max, add the key as an **API credential** instead: it stays outside the sandbox and a proxy attaches it only to requests for the hosts you list. On Team and Enterprise that feature is not available yet.

  • A cloud session reads my ~/.claude/settings.json like any local session does.
  • Plugins enabled in the repository's settings.json get installed in the cloud VM.
  • Setting network access to None stops every byte from leaving the VM.
  • Setting ANTHROPIC_API_KEY on the cloud environment switches it to API billing.
  • My GitHub token is copied into the VM so git push can work.