AWS Knowledge Base MCP Server

What this server does

One tool: `retrieve_from_aws_kb`. Takes a `query` string, a `knowledgeBaseId`, and an optional `n` (default 3) controlling how many passages come back. The server calls the Bedrock Agent Runtime's retrieve operation, which runs vector search across the documents you indexed into a Bedrock Knowledge Base and returns the top-N text passages with relevance scores. What you bring: a Bedrock Knowledge Base you have already populated via the Bedrock console or API, the KB ID, and AWS credentials with `bedrock:Retrieve` permission. The server itself does not ingest documents, manage embeddings, chunk text, or update indexes. That is Bedrock's job. This is purely the read side. Why use it: if you already pay for Bedrock and have docs indexed there (compliance runbooks, support scripts, internal wiki dumps), this server lets the model query that index from chat. The alternative is building your own RAG stack with separate embedding, vector store, and retriever components, which is more work than turning on a Bedrock KB. The credential trio (`AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, `AWS_REGION`) is passed via env vars. Long-lived IAM user keys work; for production prefer short-lived STS credentials or an instance profile, though wiring those into this stdio server takes additional glue not included.

Tools included

• retrieve_from_aws_kb

  Retrieve a passage from a Bedrock Knowledge Base by query and knowledge-base ID.

The authoritative list comes from the server's tools/list at runtime. This catalog entry may lag the upstream until the next ToolPiper release.

Install in ToolPiper

ToolPiper handles the JSON configuration. Click Install, fill in any required tokens, and the server is ready in your chat.

Install in another client

{
  "mcpServers": {
    "aws-kb-retrieval": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-aws-kb-retrieval"
      ]
    }
  }
}

Example prompts

Security notes

You are passing AWS credentials with `bedrock:Retrieve` scope in environment variables. Use a dedicated IAM user or role limited to that single permission and the specific knowledge base ARN; do not reuse admin keys. Static keys can leak via shell history, config backups, or screen shares, so prefer short-lived STS credentials where your tooling supports them. The server does not write to AWS, so the worst case is information disclosure of whatever is indexed in the KB (which can still include PII or trade secrets depending on what you ingested). The upstream Anthropic reference is archived and AWS has not published an MCP replacement, so this remains the canonical option until they do.

Related tools