AWS MCP Servers (Official)

What is AWS MCP Servers (Official)?

AWS MCP Servers (Official) is the awslabs/mcp repository: a suite of open source AWS MCP server packages that give AI assistants up-to-date AWS documentation, best practices and tools for services such as Lambda, EKS, DynamoDB and CloudWatch.

The Apache-2.0 monorepo holds more than 50 specialized servers, each installed separately with uvx. The repository README also points to the hosted AWS MCP Server, a managed remote endpoint run by AWS that combines AWS API access with documentation search. AWS now recommends its Agent Toolkit for AWS, which includes that managed server, for production work; the open source servers keep working and accepting contributions.

How to use AWS MCP Servers (Official)?

There are two ways to connect an AI client to AWS: run one of the local open source servers, or connect to the managed AWS MCP Server.

Local servers (awslabs/mcp). Prerequisites from the README:

1. Install uv from Astral.
2. Install Python with uv python install 3.10.
3. Configure AWS credentials with access to the services you need.
4. Add the server to your MCP client configuration.

The config shown on this page adds the Amazon Bedrock Knowledge Bases Retrieval server:

{
  "mcpServers": {
    "awslabs.bedrock-kb-retrieval-mcp-server": {
      "command": "uvx",
      "args": ["awslabs.bedrock-kb-retrieval-mcp-server@latest"],
      "env": {
        "AWS_PROFILE": "your-profile-name",
        "AWS_REGION": "us-east-1",
        "FASTMCP_LOG_LEVEL": "ERROR"
      }
    }
  }
}

Swap the package name for any other server, for example awslabs.aws-documentation-mcp-server or awslabs.aws-iac-mcp-server. Each server's README lists its own arguments and variables. On Windows the README uses uv tool run --from <package>@latest <package>.exe instead of uvx. Docker images are published to the public AWS ECR registry (public.ecr.aws/awslabs-mcp/...).

Managed AWS MCP Server. No local install is needed with OAuth. Attach the AWSMCPSignInOAuthAccessPolicy managed policy to your IAM user or role, then point your client at a regional endpoint such as https://aws-mcp.us-east-1.api.aws/mcp. For SigV4 authentication, AWS documents a local proxy run with uvx mcp-proxy-for-aws-cli@latest.

Claude Code

Local servers, from the awslabs README:

claude mcp add aws-docs uvx awslabs.aws-documentation-mcp-server@latest
claude mcp add aws-iac uvx awslabs.aws-iac-mcp-server@latest
claude mcp add aws-pricing uvx awslabs.aws-pricing-mcp-server@latest

Managed server with OAuth, from the AWS docs:

claude mcp add aws-mcp https://aws-mcp.us-east-1.api.aws/mcp --transport http

You can also use a .mcp.json file in your project root.

Claude Desktop

Add the managed server as a remote MCP server with the URL https://aws-mcp.us-east-1.api.aws/mcp?oauth=initialize, or paste a local uvx config like the one above into the Claude Desktop config file.

Cursor

Put the JSON config in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global). The README has one-click install buttons for Cursor. For the managed server, add the remote URL with ?oauth=initialize.

VS Code

Configure servers in VS Code settings or .vscode/mcp.json. The README has one-click VS Code install buttons for each server.

Codex

codex mcp add aws-mcp --url https://aws-mcp.us-east-1.api.aws/mcp?oauth=initialize

Key features of AWS MCP Servers (Official)

• Official AWS source: published by AWS Labs under Apache-2.0.
• Broad coverage: servers for documentation, infrastructure as code (CDK, CloudFormation), containers (EKS, ECS, Finch), serverless, databases (DynamoDB, Aurora, DocumentDB, Neptune, Redshift), caching, messaging, cost and healthcare.
• Pick only what you need: every server is a separate PyPI package, so you load the tools relevant to your project.
• Local stdio transport: the open source servers support stdio only; SSE support was removed on May 26, 2025.
• Managed option: the hosted AWS MCP Server adds IAM-based permissions, CloudTrail audit logging and eight regional endpoints.
• Client guides: setup steps for Kiro, Cline, Cursor, Windsurf, VS Code, Claude Code and fx, including ~/.codeium/windsurf/mcp_config.json for Windsurf.

AWS MCP Servers (Official) tools

Tools depend on which server you install. Examples from the READMEs:

• AWS Documentation MCP Server: read_documentation, search_documentation, read_sections, search_table and recommend for finding and reading AWS docs.
• AWS API MCP Server: call_aws runs validated AWS CLI commands and suggest_aws_commands proposes commands for a natural language request. Its README now says it is superseded by the managed AWS MCP Server.
• AWS Knowledge MCP Server (remote): search_documentation, read_documentation, list_regions, get_regional_availability and retrieve_skill.
• Managed AWS MCP Server: knowledge tools such as aws___search_documentation and aws___retrieve_skill, plus API tools aws___run_script (Python in a sandbox with AWS API access), aws___get_presigned_url and aws___get_tasks.

Use cases of AWS MCP Servers (Official)

• Current AWS answers: "Using the AWS Documentation MCP Server, explain how to configure Bedrock inline agents."
• Infrastructure as code: generate CDK or CloudFormation that follows current best practices with the AWS IaC MCP Server.
• Cost estimates: "What would be the estimated monthly cost for this CDK project?" with the AWS Pricing MCP Server.
• Operations and troubleshooting: analyze CloudWatch metrics, alarms and logs, or manage EKS and ECS deployments.
• Database work: query Aurora PostgreSQL or MySQL through the RDS Data API, or get DynamoDB data modeling guidance.
• Account-wide tasks: list resources and check their settings across services with the managed server's aws___run_script.

AWS MCP Servers (Official) vs alternatives

• Managed AWS MCP Server vs local awslabs servers: the managed server needs no local process with OAuth and logs calls in CloudTrail. Local servers run on your machine, keep credentials on your machine and cover specialized services such as databases, containers and healthcare. AWS recommends removing aws-api-mcp-server and aws-knowledge-mcp-server when you switch, to avoid duplicate tools.
• AWS MCP (RafalWilinski): a community Node.js server for Claude Desktop that queries your AWS environment with local credentials. It is not affiliated with AWS.
• AWS MCP Server (alexei-led): a community server that runs AWS CLI commands in a Docker container.

Troubleshooting AWS MCP Servers (Official)

• Server fails to start or reports missing arguments: run it by hand to see the error, for example timeout 15s uv tool run <MCP Name> <args>.
• Slow startup: the @latest suffix checks PyPI on every launch. Remove it and manage the cache with uv cache clean <tool> if load time matters.
• env values ignored in Docker: pass variables with --env, --env-file and --volume, because the "env" block does not reach the container.
• ExpiredTokenException with the managed server: refresh credentials (aws login or aws sso login) and restart the client.
• 400 error after OAuth sign-in: attach the AWSMCPSignInOAuthAccessPolicy managed policy to your IAM principal.