From model discovery to deployment with Kiro Power and Claude Code
Automate the full AI model lifecycle on AWS Neuron using natural language with Kiro Power and Claude Code. Covers model discovery, instance recommendation, training, and deployment on Amazon SageMaker AI, powered by MCP servers and Amazon Bedrock.
AWS Neuron delivers optimized price-performance and developer simplicity for AI model training and inference on purpose-built chips. This article introduces two solutions — a Kiro Power and a Claude Code Skill — that bring this capability to developers through natural language, automating the full AI model lifecycle on Amazon SageMaker AI using MCP servers.
What we built
Two MCP-powered solutions with 11 tools covering the full AI model lifecycle:
Model discovery: Integrates with any MCP-compatible model source (for example, Hugging Face Hub) to search, discover, and evaluate models.
Deployment and training: A custom SageMaker Neuron MCP server that handles instance recommendation, training, deployment, and inference — all automated.
Available as:
- Kiro Power — for developers using Kiro IDE (GUI-based)
- Claude Code Skill — for developers using Claude Code CLI (terminal-based), powered by Amazon Bedrock
The architecture is modular — model discovery integrates with any MCP-compatible server, while the AWS Neuron deployment layer stays the same and flags compatibility considerations before resources are created.
Architecture
Figure 1: Kiro Power architecture
Figure 2: Claude Code Skill architecture
The architecture chains two MCP servers (Figures 1 and 2) — one for model discovery (e.g., Hugging Face Hub) and one for deployment, training, inference, and instance recommendations on Neuron (11 tools). Anthropic Claude on Amazon Bedrock acts as an intelligent layer that orchestrates the tools and validates model responses for accuracy.
The key difference between the two solutions is the client interface:
- Kiro Power bundles steering docs and MCP servers inside a single
power.jsonmanifest, loaded by Kiro IDE. - Claude Code Skill uses
SKILL.mdfor context and configures MCP servers separately viamcp-config.json, running in the terminal.
When to use which
| Kiro Power | Claude Code Skill | |
|---|---|---|
| Interface | GUI (Kiro IDE) | Terminal (CLI) |
| Best for | Visual workflows, integrated development | CLI workflows, scripting, automation |
| Config | power.json (bundled) | SKILL.md + mcp-config.json |
| Backend | Anthropic Claude on Bedrock | Anthropic Claude on Bedrock |
| MCP tools | 11 (identical) | 11 (identical) |
| Routing | Automatic (tool-driven) | Explicit (user-driven) |
Get started
The full implementation is available on aws-samples on GitHub:
Kiro Power: GitHub
Claude Code Skill: GitHub
Clone the repo of your choice, configure your MCP servers, and start deploying models through natural language on AWS Neuron.
What’s next
- Native PyTorch backend support for Neuron (TorchNeuron)
- Additional model discovery integrations
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article