The short answer: RareCloud ships an official MCP server, @rarecloudio/mcp-server 0.4.0, with 222 tools: 110 read tools and 112 write tools. You add it to Claude Code, Cursor or any MCP-compatible client with one command or one config block and a scoped API token, and your agent can work with your catalog, servers, invoices and credit balance immediately. The token's scopes decide how far it can go: hand it read scopes to inspect, add write scopes when you want it to deploy and manage infrastructure.

What is MCP and why does a cloud provider ship one?

The Model Context Protocol (MCP) is the open standard AI assistants use to talk to external tools. Instead of pasting dashboards into a chat, your agent queries live data itself: which plans exist, what a server costs per hour, whether an invoice is paid.

We ship one because the people we build for increasingly do not click dashboards. They tell an agent what they want, and the agent needs a safe, structured way to act.

How do I connect Claude Code or Cursor?

Create an API token in the console: Account, then API tokens. Scope it to what you want the agent to do. Read scopes are a safe place to start.
In Claude Code, add it with one command (options go before the server name, the launch command after --):

claude mcp add rarecloud -e RARECLOUD_API_TOKEN=rc_pat_... -- npx -y @rarecloudio/mcp-server

In Cursor, Claude Desktop or any other MCP client, add it to the client config:

{
  "mcpServers": {
    "rarecloud": {
      "command": "npx",
      "args": ["-y", "@rarecloudio/mcp-server"],
      "env": { "RARECLOUD_API_TOKEN": "rc_pat_..." }
    }
  }
}

Ask your agent something real: "what cloud plans does RareCloud offer under 10 euro?" or "is my last invoice paid?"

What can the MCP server do?

The MCP server gives your agent live access to your account: catalog, services, billing and account state, and the operations that manage the infrastructure behind them. What an agent can actually do is decided by the API token you connect it with. A read-scoped token lets it look but not touch; a token with write scopes lets it deploy and change things. That is the safety model: an agent only ever holds the permissions you granted its token, so a prompt-injected agent can embarrass you but never exceed the scopes you set.

On top of the scopes, tools that spend money, delete something, interrupt something running or grant access refuse to run unless the call passes confirm: true, so your agent has to come back to you first. Some operations have no tool at all, by design: password and 2FA changes, API token and payment method management, credit top-ups, invoice payment, and the switch that makes a resource read-only.

That switch is yours alone. In the console you can turn API access off for any single server, cluster, volume or domain: the agent can still see it, but every change to it is refused. Use it for the server that runs your production database.

How does a deploy actually happen?

Under the hood, write actions reach the REST API with a token scoped to allow them:

curl -X POST https://api.rarecloud.io/v1/services \
  -H "Authorization: Bearer rc_pat_..." \
  -H "Content-Type: application/json" \
  -d '{"category":"cloud-vm","productId":"g-2vcpu-8gb","region":"bucharest-ro","imageId":"ubuntu-24.04","sshPublicKey":"ssh-ed25519 ..."}'

Your coding agent can generate and run this call like any other shell command. The token's scopes are the contract: a token with services:read alone can never place that order. Before ordering, the agent can send the same body to POST /v1/services/preflight (the MCP tool check_order) to learn whether the order would be accepted, without creating or charging anything. To make a retry safe, add an Idempotency-Key header with a fresh value and reuse it if the call has to be repeated: the API places the order once. The MCP server does this for you. Service and billing actions show up in your account activity log, so you can see what your agent did.

What does this cost to try?

The smallest cloud server is €3.99 per month, billed hourly with an automatic monthly cap. Running one for an afternoon of agent experiments costs a few cents.

Frequently asked questions

Does the agent need my account password? No. Tokens are separate credentials with their own scopes, and you can revoke each one independently.

Can I limit what the agent spends? Yes. Tools that spend money or delete something refuse to run unless the call passes an explicit confirm, scopes decide what the token can touch, and your account limits cap how much can run.

Which agents work? Anything that speaks MCP (Claude Code, Claude Desktop, Cursor and others) plus anything that can call a REST API, which is every coding agent we know of.

Deploy a VPS with your coding agent: the RareCloud MCP setup