Cloud World Model Now Has a Streaming MCP Server

Starting today, paste this URL into any MCP-compatible AI client:

https://www.cloudworldmodel.ai/mcp

No npx install. No Docker container. No local process. You're immediately connected to the simulation engine over Streamable HTTP — the stateless MCP transport that works through any firewall, from any IDE, over a single HTTP endpoint.

Anonymous vs. authenticated

The server has two tiers. You can evaluate it without creating an account.

The anonymous tier is real — not a stub. It runs actual simulations against the full engine, with limits that let you evaluate tool quality before committing to a key.

Cursor and Windsurf config

For clients that support remote MCP server URLs with custom headers, add this to your config:

{
  "mcpServers": {
    "cloud-world-model": {
      "url": "https://www.cloudworldmodel.ai/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

To connect in anonymous demo mode, omit the headers field entirely.

One-click via Smithery

The server is listed on Smithery. Click Connect, optionally paste your API key, and Smithery handles the routing automatically — no config editing required.

smithery.ai/servers/canvascloudai/cloud-world-model →

Three things it's good for right now

Pre-deploy validation

Describe your architecture in plain language — “3-tier app, 10k RPS, 4 EC2 instances behind an ALB, RDS Multi-AZ” — and the simulation runs a 30-step model of CPU saturation, autoscaling behavior, database connection pressure, and hourly cost. The question “will this design hold under load?” has a concrete answer before you write a single line of Terraform.

The simulation engine matches real provider behavior at 97%+ accuracy across AWS, GCP, Azure, OCI, and DigitalOcean — same latency curves, autoscaling delays, and cost models the provider actually exhibits under load.

Chaos testing without a staging environment

Inject a database crash. Measure how long recovery takes. Get a resilience score. Run the same test across five architectures and compare. The full chaos suite runs in seconds with no blast radius on real systems.

This is the practical path for teams without a proper staging environment — or for pre-merge validation of infrastructure changes that would otherwise require a test cluster.

RL agent training for cloud optimization

The MCP server exposes the full RL environment API. An agent can create an environment, step through episodes, collect rewards based on cost/latency/uptime trade-offs, and iterate on a policy — entirely through MCP tool calls.

Each RL step models realistic provider behavior including autoscaling delays, cold-start penalties, and spot interruption probabilities. It's the zero-cost way to train a cloud optimization policy before deploying it against real infrastructure.

What's in the tool set

The 49 authenticated tools cover simulation lifecycle, hybrid ML + rule-based stepping, traffic pattern injection, failure injection, multi-cloud comparison, AI-generated analysis, RL environment management, and async job management. The 6 demo tools cover the core loop: create, step, inspect, and explain.

Full tool reference and step-by-step setup:

MCP Quick Start guide →Full agent API reference →Get an API key →