Qudata MCP

What is Qudata MCP?

Qudata MCP is a Model Context Protocol server that gives AI agents typed access to Qudata.ai: the GPU marketplace, GPU catalog, launch templates, and instance management. With MCP, an agent can search GPU offers, check prices and availability, prepare templates, and perform account actions with the user's API key.

The MCP server runs over Streamable HTTP and is available at:

https://api.qudata.ai/mcp

Use this page when you want to connect Qudata to Cursor, Claude Desktop, or another MCP client and manage GPU infrastructure with an AI agent.

Connect a client

For Cursor, Claude Desktop, and other MCP clients, you can use mcp-remote. Add the qudata server to your MCP client configuration:

{
  "mcpServers": {
    "qudata": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.qudata.ai/mcp",
        "--header",
        "Authorization: Bearer YOUR_QUDATA_API_KEY"
      ]
    }
  }
}

You can generate an API key on the API keys page. For MCP clients, use the Authorization: Bearer <key> header. The shared backend header X-API-Key is also supported.

Discovery for AI agents

Qudata publishes machine-readable endpoints that help AI systems and search agents understand the service capabilities:

• /.well-known/mcp.json — MCP transport, authentication, and scopes.
• /.well-known/ai-plugin.json — legacy manifest for compatibility.
• /llms.txt — short Qudata MCP description for LLMs.
• /llms-full.txt — extended tools, parameters, and safety limitations.

These files help LLMs and agent platforms index Qudata MCP correctly, discover available tools, and respect safety rules before taking actions in an account.

MCP tool capabilities

Marketplace

search_market_offers searches GPU offers in the Qudata marketplace. It supports filters by GPU, provider, region, availability, price, currency, and locale. Use it for prompts like: "find the cheapest RTX 4090 offers available for rent" or "choose a GPU for model training in Europe".

get_offer returns an extended offer card by offer_id: GPU, prices, rentable, in_stock, run type, and other details needed before renting.

GPU catalog

list_gpus returns GPU models with filters by provider, marketplace availability, and search query.

get_gpu finds a GPU by ObjectId, translit_name, or fuzzy name.

gpu_profitability calculates GPU profitability for a period. If hourly_price is not provided, Qudata uses the market price.

Templates

list_templates lists public and user launch templates: instance, vm, and agent.

get_template returns a template visible to the API key owner.

create_template creates a Docker template of type instance. You can define the image, tag, environment variables, launch command, ports, SSH, Web UI, minimum disk size, supported GPUs, and template privacy.

update_template updates a user Docker template.

update_vm_template updates a user VM template without changing the VM image.

clone_vm_template copies a visible VM template into the user's templates.

delete_template deletes a user template and requires confirm=true.

Instances

list_instances returns instances owned by the API key owner.

rent_instance creates a paid GPU instance from an offer and a template. By default, the tool returns a preview and does not create a resource without explicit confirm=true.

Before creating an instance, Qudata checks that:

• the offer exists;
• the template is visible to the API key owner;
• the offer is rentable: rentable == true;
• the server is in stock: in_stock == true;
• the offer run type is compatible with the template type.

Compatibility rules:

• RunType.vm requires TemplateType.vm;
• RunType.instance requires TemplateType.instance.

manage_instance controls a user instance: start, pause, reboot.

destroy_instance deletes an instance owned by the API key owner and requires confirm=true.

attach_ssh_key and detach_ssh_key attach or detach an SSH key from an instance.

Safety and confirmations

Read-only tools for GPU search and marketplace analysis can be used to compare configurations. Tools that change account state or create paid resources require an API key and explicit confirmations.

Important cases:

• rent_instance can charge the balance and create an external provider resource;
• a repeated confirmed rent_instance call can create another instance;
• destroy_instance terminates a resource and starts deletion finalization;
• delete_template removes a user template.

For destructive and paid tools, use a two-step flow: ask the agent to get a preview without confirm, review the parameters, and only then repeat the action with confirm=true.

Example agent prompts

Show 5 cheapest RTX 4090 offers available for rent.

Create a private template for docker image ghcr.io/company/worker:latest with SSH and UI on port 7860.

Check the offer and template for renting, but do not create an instance without confirmation.

Show my active instances and reboot the selected instance_id.

How to write better prompts

For best results, include the GPU model, budget, region, rental period, currency, and template requirements. For example: "find an available NVIDIA H100 in Europe, sort by price, check compatibility with a VM template, and show a rent preview without confirmation".