AWS MCP Server

What AWS MCP Server MCP server does

AWS MCP Server MCP server provides a bridge between MCP clients and MCP servers on AWS that require AWS IAM authentication rather than OAuth. It does not define a fixed collection of AWS service tools. Instead, it forwards requests to an MCP endpoint and makes that endpoint's tools available to the connected client.

The package has two distributions for different uses. mcp-proxy-for-aws-cli is the version-pinned command-line distribution intended for running the proxy with uvx. mcp-proxy-for-aws is the library distribution for importing mcp_proxy_for_aws into Python applications and connecting agent frameworks such as LangChain, Strands Agents, and LlamaIndex.

How it works

The AWS MCP Server MCP server uses AWS credentials available through the local AWS CLI configuration, environment variables, or IAM roles. It signs MCP requests with AWS Signature Version 4 before sending them to the configured endpoint. This avoids requiring each MCP client or application to implement its own SigV4 request signing.

As a proxy, it runs locally between an MCP client and an upstream AWS MCP server. Supported clients named in the documentation include Claude Desktop and Kiro CLI. The proxy can discover tools from the upstream server dynamically. As a library, it allows application code to manage MCP sessions and connect AWS IAM-secured endpoints from supported agent frameworks.

Setup and configuration

The prerequisites are Python 3.10 or newer, the uv package manager, and configured AWS credentials. Docker Desktop is optional for users running the published container image.

The proxy requires an MCP endpoint URL. The AWS service used for signing can be supplied with --service; otherwise, the proxy attempts to infer it from the endpoint. --region selects the AWS region and falls back to AWS_REGION when omitted. AWS profiles can be supplied with --profile; the first profile is the default, and additional profiles can support per-call switching through the aws_profile tool parameter.

Additional options control injected key-value metadata, retries, logging, and request timeouts. The --read-only option disables tools that may require write permissions, based on the upstream tools' readOnlyHint annotations.

Tools and capabilities

The AWS MCP Server MCP server supports:

• SigV4 signing of requests to IAM-protected MCP endpoints
• Dynamic discovery of tools exposed by an upstream MCP server
• Multiple AWS profiles and per-call profile selection
• AWS region selection and metadata injection
• Optional read-only operation
• Configurable retries, timeout, and log level
• Local execution through uvx, a checked-out repository, or Docker
• Programmatic use from Python agent applications

The CLI distribution can be started with uvx mcp-proxy-for-aws-cli@latest followed by the upstream SigV4 MCP endpoint URL. A local checkout can run the proxy with uv run mcp_proxy_for_aws/server.py and the endpoint URL. Docker images are published through the public AWS ECR registry.

Limitations and notes

An upstream MCP endpoint is required; the proxy is not an independent catalog of AWS operations. The client must be able to use a local MCP bridge, and the AWS credentials used by the proxy must already be configured through one of the supported credential sources.

The library and CLI distributions serve different dependency-management goals. The CLI package pins its dependency tree to tested versions, while the library package uses loose, co-resolvable dependency ranges for integration into an application's environment. The first uvx run may take tens of seconds while dependencies download and cache; later runs are expected to start faster.