Kubeflow MCP Server

AI-powered interface for Kubeflow Training via Model Context Protocol (MCP). Enable your AI assistants to manage distributed training jobs, fine-tune LLMs, and monitor workloads on Kubernetes — all through natural language.

Note: This project is in early development. APIs may change between versions.

Table of Contents

• Overview

• Quick Start

  • Installation

  • Configuration

• Tools

• Prompts

• CLI

• Local Agent

• Development

• Roadmap

• Contributing

• Community

• License

Related MCP server: k8s-mcp-server

Overview

The Kubeflow MCP Server bridges AI assistants/agents with Kubeflow's training infrastructure. Instead of writing YAML manifests or learning Kubernetes APIs, simply describe what you want to train and let AI handle the complexity.

Key Benefits

• Natural Language Interface: Describe training jobs in plain English — "fine-tune Llama-3 on my dataset with 4 GPUs"

• Smart Resource Planning: AI estimates GPU/memory requirements before job submission

• Real-time Monitoring: Stream logs, track progress, and debug failures conversationally

• Safe by Design: Preview configurations before submission, built-in validation and guardrails

• Multi-Client Support: Works with Claude Desktop, Cursor IDE, MCP Inspector, or custom agents

Compatibility

This MCP server wraps the Kubeflow Training SDK TrainerClient API. All training operations (fine-tuning, custom scripts, container jobs) use SDK types like BuiltinTrainer, CustomTrainer, TorchTuneConfig, and LoraConfig.

Quick Start

Installation

pip install kubeflow-mcp

Configuration

Add to your MCP client configuration:

{
  "mcpServers": {
    "kubeflow": {
      "command": "kubeflow-mcp",
      "args": ["serve", "--persona", "ml-engineer"]
    }
  }
}

{
  "mcpServers": {
    "kubeflow": {
      "command": "podman",
      "args": [
        "run", "--rm", "-i",
        "-v", "${HOME}/.kube:/home/mcp/.kube:ro",
        "ghcr.io/kubeflow/mcp-server:latest",
        "serve"
      ]
    }
  }
}

Replace podman with docker if needed.

npx @modelcontextprotocol/inspector uv run kubeflow-mcp serve

Try It

"What training jobs are running in my cluster?"
"Fine-tune google/gemma-2b on squad dataset with 2 GPUs"
"Show me logs for the failed training job"

Tools

fine_tune(
    model="google/gemma-2b",
    dataset="squad",
    num_nodes=2,
    gpu_per_node=1,
    confirmed=True
)

Ask: "How much GPU memory do I need for Llama-3-70B?"

{
  "model": "meta-llama/Llama-3-70B",
  "parameters": "70B",
  "estimated_memory_gb": 140,
  "recommended_gpus": 4,
  "gpu_type": "A100-80GB"
}

Prompts

MCP prompts provide structured guidance for common workflows. MCP clients can discover and use these prompts:

MCP clients that support prompts (like Claude Desktop) can list and invoke these prompts directly. The prompts provide detailed, parameterized guidance that helps ensure successful training operations.

Example with parameters:

fine_tuning_workflow(model="meta-llama/Llama-3.2-3B", dataset="tatsu-lab/alpaca")

Resources

MCP resources provide read-only reference data that clients can fetch without consuming tool calls:

CLI

# Server
kubeflow-mcp serve                              # Start MCP server
kubeflow-mcp serve --clients trainer            # Specify client
kubeflow-mcp serve --persona ml-engineer        # Set persona
kubeflow-mcp status                             # Show server status

# Agent
kubeflow-mcp agent --backend ollama --model qwen3:8b
kubeflow-mcp agent --backend ollama --mode progressive
kubeflow-mcp agent --backend ollama --thinking  # Enable thinking output

Local Agent

Run a fully local AI agent with Ollama — no cloud APIs required:

pip install kubeflow-mcp[agents]
ollama pull qwen3:8b
kubeflow-mcp agent --backend ollama --model qwen3:8b

kubeflow-mcp agent --backend ollama                    # Full mode (default)
kubeflow-mcp agent --backend ollama --mode progressive # Hierarchical discovery
kubeflow-mcp agent --backend ollama --mode semantic    # requires sentence-transformers

Full mode connects via the standard MCP stdio protocol, identical to Cursor and Claude Desktop.

For 8K context models, use --mode progressive or --mode semantic.

Development

make dev           # Install dev dependencies
make check         # Lint + type check
make test          # Run tests
make pre-commit    # All checks before commit

See make help for all commands. For detailed setup, see docs/DEVELOPMENT.md.

Roadmap

See ROADMAP.md for details.

For MCP clients, agent frameworks, and production Kubernetes companions (Kyverno, ingress/OIDC, observability) — and what this project supports in-repo vs at deploy time — see Ecosystem and platform integrations in the roadmap.

Contributing

git clone https://github.com/kubeflow/mcp-server.git
cd mcp-server
make dev && make pre-commit

Look for good first issue labels.

Community

• Slack: #kubeflow-ml-experience on CNCF Slack

• Meetings: Kubeflow SDK and ML Experience bi-weekly

License

Apache-2.0