Things to try

Once your client is connected, here's what the American Cloud MCP server is good at. Prompts below are starting points — adapt them to your account and phrasing.

Explore and audit (read-only)

These work with read-only access: a connected app set to Read-only, or a local server with a Read-only token. They're the best way to build trust in the setup before enabling writes.

[ claude code · read-only audit ]

Take inventory:

What's running in my account right now? List every VM with its region, size, IP, and power state.

Understand your spend:

Walk through everything in my account and estimate what each resource costs per month. What's the biggest line item?

Security review:

Audit my firewall rules and network ACLs. Is anything open to the whole internet that doesn't need to be? Which ports are exposed on each public IP?

Find waste:

Are there any block storage volumes not attached to a VM, or reserved public IPs not assigned to anything?

Incident triage:

I can't reach my app on port 443. Check the VM's power state, the firewall rules on its public IP, any port forwarding rules, and whether it's behind a load balancer with healthy members.

Snapshot hygiene:

List all my snapshots with their creation dates. Which volumes have no recent snapshot?

Price before you build (read-only)

Cost estimates are tools too, so the assistant can compare options before creating anything:

What would a 3-node Kubernetes cluster with mid-size workers cost per month?

Compare the monthly cost of one large VM versus three small ones behind a load balancer.

Provision and manage (requires write access)

With the hosted server, a known app such as Claude, ChatGPT, Codex or Cursor has Full access from the start; an app that registered itself, such as Grok, starts read-only, and you choose Full access for it in Connected apps. With the local server, you need a Full access token and the --allow-writes flag (see overview). Then the assistant can build:

A development environment in one prompt:

Create a small Ubuntu VM in us-west on an isolated network, use my SSH key, and open ports 22, 80, and 443. Show me the cost estimate first.

Pre-deploy safety net:

Take a snapshot of every volume attached to my database VMs, named with today's date.

Database plus DNS:

Set up PostgreSQL on a small VM, keep it listening on the private network only, then add an A record for api.example.com pointing at my load balancer's IP.

Routine operations:

Reboot the staging VM and confirm it comes back up.

Resize the volume on my build server to 200 GB.

Code and cloud together (coding agents)

In a coding agent like Claude Code, the MCP server combines with your terminal — the assistant can provision infrastructure and use it in the same session. This is also the only way an assistant runs commands inside a running VM: the MCP server manages your account, and the agent's own shell does the SSH. A chat assistant without a shell can configure a new VM with cloud-init at creation time or open a browser console for you; see Inside a VM.

Build and deploy:

Register my public key from ~/.ssh/id_ed25519.pub with American Cloud, then create a VM for this project with that key, open ports 22 and 80, install nginx over SSH and deploy the site in ./dist.

Kubernetes workflow:

Fetch the kubeconfig for my cluster and check whether all deployments are healthy.

Infrastructure as conversation:

Stand up a staging copy of my production setup — same network layout, half the VM size — and give me a cost comparison when you're done.

Tips for better results

• Scope the tool list. On the local server, if you only work with compute and DNS, set --services compute,dns — a smaller tool list helps the assistant pick the right tool faster. The hosted server always lists every tool; a Custom permission set in Connected apps limits what the app can change, not the length of the list.
• Ask for cost estimates first. The assistant has pricing tools; make "estimate before create" part of your prompts until it's a habit.
• Start read-only. Run a week of audits and triage with read-only access before enabling writes. You'll learn what the assistant is good at with zero risk.
• Be specific about regions and sizes when provisioning, or ask the assistant to list the options and recommend one.