Skymcp

SkyMCP — SkyPilot ML Training Ecosystem Plugin

End-to-end ML training platform for Claude Code. Launch, monitor, fix, iterate, ablate, and serve models on any cloud GPU using SkyPilot with production-grade frameworks.

What It Does

Once installed, this plugin transforms Claude Code into a senior ML engineer + MLOps specialist + cloud cost optimizer:

• Launch training jobs on any cloud GPU (AWS, GCP, Azure, Lambda, RunPod, 20+ providers)
• Pick the right framework automatically (NeMo, Axolotl, torchtune, TRL, DeepSpeed, Megatron)
• Monitor training with W&B/TensorBoard integration, automatic issue detection
• Diagnose and fix OOM, NaN, loss plateaus, gradient explosion, slow throughput
• Run experiments — hyperparameter sweeps, architecture ablations, scaling law studies
• Optimize costs — spot instances, multi-cloud failover, autostop, budget management
• Evaluate models with lm-evaluation-harness across standard benchmarks
• Deploy models with vLLM on SkyServe with autoscaling
• Real data pipelines — NeMo Curator, FineWeb, DCLM patterns for production data curation

Prerequisites

• SkyPilot installed and configured (pip install skypilot[aws,gcp])
• Cloud credentials configured (sky check)
• Node.js 18+ (for MCP server and hooks)

Installation

# Test locally
claude --plugin-dir /path/to/skymcp

# Or copy to your project
cp -r skymcp/.claude-plugin your-project/

Components

Slash Commands (9)

Command	Description
/sky-launch	Generate YAML, validate, estimate cost, launch training
/sky-status	Dashboard of all clusters, jobs, services, costs
/sky-logs	Stream and analyze logs, detect issues
/sky-down	Safe teardown with cost savings report
/sky-cost	Spending analysis and optimization suggestions
/sky-sweep	Launch hyperparameter sweep across cloud GPUs
/sky-eval	Run model evaluation benchmarks
/sky-serve	Deploy model with vLLM + SkyServe autoscaling
/sky-recipe	Generate end-to-end pipeline (data + train + eval + serve)

Auto-Activating Skills (8)

Activate automatically based on conversation context:

• skypilot-core — CLI reference, YAML spec, env vars, 21 gotchas
• ml-training-frameworks — NeMo vs Axolotl vs torchtune vs TRL decision matrix
• data-pipeline-design — NeMo Curator, FineWeb, dedup, quality filtering
• training-monitoring — W&B, TensorBoard, OOM/NaN/plateau diagnosis
• model-evaluation — lm-eval-harness, lighteval, benchmark selection
• cost-optimization — Spot strategies, multi-cloud failover, budget management
• checkpoint-management — Distributed checkpoints, LoRA merging, GGUF conversion
• distributed-training — Multi-node, DeepSpeed ZeRO, FSDP2, InfiniBand

Agents (5)

Agent	Role
training-orchestrator	Full lifecycle: framework selection, launch, monitor, recover, iterate
experiment-scientist	Ablation design, scaling laws, sweep comparison
config-validator	YAML validation, gotcha detection, cost estimation
training-doctor	Diagnose OOM, NaN, plateau, divergence, slow throughput
cloud-optimizer	Spending analysis, spot migration, savings recommendations

Hooks (3)

• PostToolUse — Captures sky launch output, tracks job IDs
• SessionStart — Loads active clusters/jobs/costs into context
• PreToolUse — Validates destructive ops, suggests spot/autostop

MCP Server (7 tools)

sky_status, sky_launch, sky_logs, sky_down, sky_cost, sky_gpus, sky_check

Recipe Templates (7)

Ready-to-use SkyPilot YAML recipes in references/recipes/:

• nemo-pretraining.yaml — Multi-node NeMo 2.0 on H100 cluster
• axolotl-finetune.yaml — QLoRA fine-tuning with Axolotl
• torchtune-finetune.yaml — Full fine-tuning with torch.compile
• trl-dpo.yaml — DPO preference alignment
• vllm-serve.yaml — Production inference with SkyServe
• nemo-curator.yaml — GPU-accelerated data curation
• full-pipeline.yaml — 4-stage: data prep + train + eval + serve

Quick Start

# Check cloud credentials
sky check

# Launch a fine-tuning job
# (use /sky-launch in Claude Code for interactive workflow)
sky jobs launch references/recipes/axolotl-finetune.yaml \
  --env HF_TOKEN=$HF_TOKEN \
  --env WANDB_API_KEY=$WANDB_API_KEY

# Monitor
sky jobs queue
sky jobs logs JOB_ID

# Evaluate
sky jobs launch references/recipes/eval.yaml

# Serve
sky serve up references/recipes/vllm-serve.yaml -n my-model

Architecture

skymcp/
├── .claude-plugin/plugin.json    # Plugin manifest + MCP config
├── skills/                        # 17 skills (8 auto + 9 commands)
│   ├── skypilot-core/            # Core SkyPilot reference
│   ├── ml-training-frameworks/   # Framework selection
│   ├── sky-launch/               # /sky-launch command
│   └── ...
├── agents/                        # 5 autonomous agents
├── hooks/                         # 3 event hooks + scripts
├── mcp/                          # MCP server (TypeScript)
└── references/recipes/           # 7 ready-to-use YAML templates

License

MIT