Skip to content

Add Amazon Bedrock Guided MCP Power #82

@dgallitelli

Description

@dgallitelli

Proposal: Add amazon-bedrock Guided MCP Power

What

A Guided MCP Power for building generative AI applications on Amazon Bedrock, integrating two MCP servers from awslabs/mcp.

Coverage

  • Model catalog — All foundation model IDs and CRIS inference profile IDs (pulled from live AWS APIs), provider reference, embedding model selection
  • Converse API — Unified model invocation, streaming, tool use, multimodal inputs, prompt caching, extended thinking/reasoning
  • Agents — Action groups (function definitions, OpenAPI, return control), Lambda handlers, code interpreter, KB association
  • Knowledge Bases (RAG) — Vector stores (OpenSearch Serverless, Aurora, S3 Vectors), data sources, chunking strategies, retrieve and generate
  • Guardrails — Content filters, PII protection, topic denial, word blocking, contextual grounding checks
  • Model customization — Supervised fine-tuning, reinforcement fine-tuning, distillation, custom model import
  • Prompt management and Flows — Reusable versioned prompts with variables, multi-step AI workflow orchestration

MCP Servers

  • awslabs.bedrock-kb-retrieval-mcp-server — Knowledge Base querying with reranking
  • awslabs.aws-bedrock-custom-model-import-mcp-server — Custom model lifecycle management

Note: Deliberately excludes amazon-bedrock-agentcore-mcp-server as it already has its own aws-agentcore power.

Structure

amazon-bedrock/
├── POWER.md
├── mcp.json
└── steering/
    ├── model-catalog.md
    ├── converse-api.md
    ├── agents.md
    ├── knowledge-bases.md
    ├── guardrails.md
    ├── model-customization.md
    └── prompts-and-flows.md

Why

Bedrock has a broad API surface across multiple boto3 clients (bedrock, bedrock-runtime, bedrock-agent, bedrock-agent-runtime). Without explicit steering, AI agents commonly:

  • Use invoke_model() with model-specific JSON payloads instead of the Converse API
  • Use single-region model IDs instead of CRIS inference profiles (lower throughput)
  • Mix up build-time vs runtime clients
  • Use outdated model IDs

The always-on guardrails table in POWER.md prevents these mistakes regardless of which steering file is loaded.

PR

See the linked PR below.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions