Agent Builder MCP Server

An MCP server that provides tools and knowledge for building, deploying, and managing agents on AWS Transform.

This server works with any MCP-compatible client — Kiro, Claude Code, Cursor, Windsurf, or any other IDE/tool that supports the Model Context Protocol.

Installation

pip install agent-builder-mcp-aws-transform

This installs an agent-builder-mcp command that speaks MCP over stdio.

Quick start

With Kiro (recommended)

Install the AWS Transform Agent Toolkit Kiro Power for the full guided experience — it configures this MCP server automatically and adds steering rules that guide Kiro through the entire agent-building process.

With any MCP client

Add the server to your MCP configuration:

{
  "mcpServers": {
    "agent-builder": {
      "command": "uvx",
      "args": ["--from", "agent-builder-mcp-aws-transform", "--with", "mcp>=1.0.0,<2.0.0", "agent-builder-mcp"]
    }
  }
}

Restart your IDE/tool. It will now have access to the agent-builder tool set.

What it provides

Tools are grouped into six categories:

• Search — retrieve documentation and examples from the bundled AWS Transform knowledge base (BM25 keyword search, no network calls, no embeddings).
• Agent registry — look up, register, update, deregister, and version agents.
• Skill operations — manage the skills an agent exposes. Deprecated: these tools will be removed in a future release; their descriptions and results carry the deprecation notice.
• Deployment — package and deploy agents to AWS.
• Validation — check agent manifests and configurations before deployment.
• CloudWatch — query agent logs and metrics.

Requirements

• Python 3.11+
• AWS credentials configured (standard boto3 credential chain) for deployment and CloudWatch tools. Search and validation work offline.
• Deployment & registration tools: Your AWS account must be allowlisted for the AWS Transform composability initiative. Contact your Partner Development Manager (PDM) or apply through AWS Partner Central. Search and validation tools work without allowlisting.

Development

Retrieval benchmark

The bundled BM25 search is gated by a quality benchmark (48 golden queries across 6 categories):

pip install -e .
python bench/eval_retrieval.py

CI enforces Recall@5 >= 0.70 and MRR >= 0.60. Current: Recall@5 = 0.80, MRR = 0.85.

License

Apache-2.0. See LICENSE and THIRD-PARTY-NOTICES.txt.