@yottascale/agent-native-infra

MCP server and agent skills for Yotta Platform — the GPU cloud for AI/ML workloads.

Give any AI agent the ability to provision GPUs, launch pods, deploy models, and manage infrastructure through natural language. Built on the Model Context Protocol (MCP).

What's included

| Layer | What it does | Count | |-------|-------------|-------| | Tools | CRUD operations for VMs, Pods, Serverless endpoints, Volumes, and Registry credentials | 37 | | Resources | GPU catalog with specs, pricing, and availability | 2 | | Prompts | Guided workflows for GPU selection, pod launch, and model serving | 3 | | Skills | Agent skill definitions for Claude Code and compatible agents | 3 |

Quick start

Prerequisites

• Node.js >= 18
• A Yotta Platform API key (get one here)

Use with Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "yotta": {
      "command": "npx",
      "args": ["-y", "@yottascale/agent-native-infra"],
      "env": {
        "YOTTA_API_KEY": "your-api-key"
      }
    }
  }
}

Use with Claude Code

claude mcp add yotta -- npx -y @yottascale/agent-native-infra

Set the API key in your environment:

export YOTTA_API_KEY=your-api-key

Use with Cursor, Windsurf, or any MCP-compatible client

{
  "mcpServers": {
    "yotta": {
      "command": "npx",
      "args": ["-y", "@yottascale/agent-native-infra"],
      "env": {
        "YOTTA_API_KEY": "your-api-key"
      }
    }
  }
}

Run locally (from source)

git clone https://github.com/yottalabsai/agent-native-infra
cd agent-native-infra
npm install
YOTTA_API_KEY=your-api-key npx tsx src/index.ts

Or point Claude Desktop / Claude Code at the local build:

{
  "mcpServers": {
    "yotta": {
      "command": "npx",
      "args": ["tsx", "/path/to/agent-native-infra/src/index.ts"],
      "env": { "YOTTA_API_KEY": "your-api-key" }
    }
  }
}

Test with MCP Inspector

YOTTA_API_KEY=your-api-key npx @modelcontextprotocol/inspector npx -y @yottascale/agent-native-infra

Tools

Pods

Interactive GPU instances for development, training, and batch jobs.

| Tool | Description | |------|-------------| | pod_create | Create a GPU pod with a Docker image, GPU type/count, ports, and env vars | | pod_get | Get pod details by ID | | pod_list | List pods, optionally filtered by region or status | | pod_delete | Delete a pod (irreversible) | | pod_pause | Pause a running pod (stops billing, preserves state) | | pod_resume | Resume a paused pod |

Serverless

Elastic (serverless) GPU endpoints for production inference.

| Tool | Description | |------|-------------| | serverless_create | Create a serverless endpoint (ALB, QUEUE, or CUSTOM mode) | | serverless_get | Get endpoint details by ID | | serverless_list | List all serverless endpoints, optionally filtered by status | | serverless_update | Update endpoint configuration | | serverless_delete | Delete an endpoint (irreversible) | | serverless_stop | Stop a running endpoint | | serverless_start | Start a stopped endpoint | | serverless_scale | Scale worker count up or down | | serverless_list_workers | List workers for an endpoint | | serverless_list_tasks | List tasks for a QUEUE-mode endpoint | | serverless_task_count | Get task status counts | | serverless_submit_task | Submit a task to a QUEUE-mode endpoint | | serverless_get_task | Get details of a specific task by ID | | serverless_worker_logs | Get logs from a specific worker |

Virtual Machines

Full GPU virtual machines.

| Tool | Description | |------|-------------| | vm_create | Create a GPU VM (on-demand or spot) | | vm_get | Get VM details by ID | | vm_list | List VMs (paginated) | | vm_types | List available VM/GPU types with region availability | | vm_rename | Rename a VM | | vm_terminate | Terminate a VM (irreversible) |

Volumes

Persistent and object storage for pods and VMs.

| Tool | Description | |------|-------------| | volume_create | Create a storage volume (S3, R2, CEPH, VENDOR) | | volume_list | List volumes by storage type (paginated) | | volume_get | Get volume details by ID | | volume_delete | Delete a volume (must be unmounted) | | volume_rename | Rename a volume | | volume_resize | Resize a CEPH or VENDOR volume |

Container Registry

Manage credentials for pulling private Docker images.

| Tool | Description | |------|-------------| | registry_list | List all registry credentials | | registry_get | Get a credential by ID | | registry_create | Create a new credential | | registry_update | Update a credential | | registry_delete | Delete a credential |

Resources

| URI | Description | |-----|-------------| | yotta://gpus | Full GPU catalog (all types with VRAM, pricing, regions) | | yotta://gpus/{gpuType} | Individual GPU type details |

Available GPUs

| GPU | VRAM | |-----|------| | NVIDIA RTX 4090 | 24 GB | | NVIDIA RTX 5090 | 32 GB | | NVIDIA A100 | 80 GB | | NVIDIA H100 | 80 GB | | NVIDIA H200 | 141 GB | | NVIDIA B200 | 192 GB | | NVIDIA B300 | 288 GB | | NVIDIA RTX PRO 6000 | 96 GB |

Prompts

gpu-selector

Interactive GPU recommendation based on model size, task type, budget, and quantization. Estimates VRAM requirements and suggests optimal configurations.

Task: fine-tuning | Model: Llama-3-70B | Budget: medium | Quantization: int4
→ Recommends H100 80GB x1 with QLoRA

launch-pod

Configure and launch a GPU pod from preset templates:

• pytorch — General deep learning (training, fine-tuning, research)
• unsloth — Fast LoRA/QLoRA fine-tuning (2-5x speedup)
• skyrl — Reinforcement learning (RLHF, PPO, GRPO)
• comfyui — Image generation (Stable Diffusion, SDXL, Flux)

serve-model

Deploy a model for inference. Supports multiple serving frameworks (vLLM, TGI, Triton) and deployment modes:

| Mode | Description | |------|-------------| | POD | Single GPU instance via pod_create — good for dev/testing | | ALB | HTTP load balancer via serverless_create — real-time inference at scale | | QUEUE | Async job queue — batch/long-running jobs | | CUSTOM | Raw container — gRPC or custom protocols |

Agent Skills

The skills/yotta-agent-skills/SKILL.md file provides structured knowledge for AI agents, including:

• VRAM estimation heuristics for sizing GPUs to models
• Template-to-image mapping for quick pod launches
• Serving framework selection guidance
• Step-by-step configuration workflows

Compatible with Claude Code and any agent framework that supports skill files.

Configuration

| Environment Variable | Required | Default | Description | |---------------------|----------|---------|-------------| | YOTTA_API_KEY | Yes | — | Yotta Platform API key | | YOTTA_API_BASE_URL | No | https://api.yottalabs.ai | API base URL |

Development

npm run dev          # Watch mode with hot reload
npm test             # Run tests
npm run test:watch   # Watch mode tests
npm run lint         # Type check
npm run build        # Compile TypeScript

Project structure

src/
├── index.ts              # Server entry point
├── config.ts             # Environment configuration
├── api/
│   ├── client.ts         # HTTP client for Yotta V2 API
│   └── types.ts          # TypeScript interfaces
├── tools/
│   ├── vms.ts            # VM tools (6)
│   ├── pods.ts           # Pod tools (6)
│   ├── serverless.ts     # Serverless tools (14)
│   ├── volumes.ts        # Volume tools (6)
│   └── registry.ts       # Registry tools (5)
├── resources/
│   ├── index.ts          # GPU catalog resources
│   └── gpus.json         # GPU type definitions
└── prompts/
    ├── gpu-selector.ts   # GPU recommendation prompt
    ├── launch-pod.ts     # Pod launch prompt
    └── serve-model.ts    # Model serving prompt
skills/
└── yotta-agent-skills/
    └── SKILL.md          # Agent skill definitions

License

MIT — see LICENSE.