MCP Server

Connect an AI agent to your CubePath infrastructure with the Model Context Protocol (MCP). The server is hosted by us, so there is nothing to install or run locally: point your MCP client at the endpoint and authenticate with an API token.

https://api.cubepath.com/mcp

Connect a client

The server speaks JSON-RPC 2.0 over streamable HTTP (one POST per message), so any MCP client with HTTP transport support works. Most of them take a configuration like this:

{
  "mcpServers": {
    "cubepath": {
      "url": "https://api.cubepath.com/mcp",
      "headers": {
        "X-API-Key": "YOUR_API_TOKEN"
      }
    }
  }
}

Authorization: Bearer YOUR_API_TOKEN works as well. Create the token from Organization → Tokens, as described in API.

Note

The endpoint accepts API tokens only. Session cookies are rejected on purpose, so no logged-in browser can drive it.

What your agent can do

150 tools cover the public surface of the API, grouped by product:

Compute — VPS, baremetal, Kubernetes, availability groups.
Networking — load balancers, DNS, floating IPs, NAT gateways, private networks, tunnels, firewalls.
Delivery and data — CDN, managed databases, transcoder.
Platform — projects, SSH keys, alerts, status, DDoS.

Tools are named after what they do: cubepath_vps_create, cubepath_k8s_cluster_list, cubepath_dns_record_create.

Scopes decide what the agent sees

tools/list returns only the tools your token's scopes allow, so a read-only token exposes read-only tools and nothing else. Give each agent a token with the minimum scopes it needs.

Every tool is annotated as read-only or destructive, which lets your client ask you to confirm before something like cubepath_vps_destroy runs.

Long-running operations

Deploying a VPS or destroying a cluster is queued, and the tool returns straight away. Ask for the resource status to follow progress instead of calling the tool again: a second call queues a second operation.

Limits

120 requests per minute per IP address on the MCP endpoint.
Request body up to 256 KB, and batches of up to 20 messages.
Tool results are capped at roughly 15 KB. Long listings are trimmed and the tool states what it left out, so a single response cannot fill the model's context.
GET returns 405. The server is stateless: no SSE, no sessions.

On this page