AWS MCP Server

What it does

The AWS MCP Server is a managed, remote MCP server run by AWS that lets AI agents search current AWS documentation, retrieve curated AWS skills and, once authenticated, call AWS APIs through your own IAM credentials. It is the server at the center of the Agent Toolkit for AWS, and it works with Claude Code, Cursor, Codex, Kiro and other MCP clients. Documentation search works without credentials; API calls and scripts need IAM authentication. AWS announced general availability of the server on May 6, 2026.

Key features and tools

AWS documents these tools in "Understanding the MCP Server tools":

AWS knowledge tools

• aws___search_documentation: searches all AWS documentation, including API references, best practices, service guides and skills (formerly Agent SOPs).
• aws___read_documentation: retrieves an AWS documentation page and converts it to markdown.
• aws___retrieve_skill: returns the full content of a domain skill, with workflows, best practices and step-by-step procedures.
• aws___list_regions: lists all AWS Regions with identifiers and names.
• aws___get_regional_availability: checks Regional availability for services, features, SDK APIs and CloudFormation resources.

AWS API tools (authentication required)

• aws___run_script: runs Python code in a sandboxed environment with AWS API access, for listing resources, parallel calls and multi-step workflows.
• aws___get_presigned_url: generates pre-signed Amazon S3 URLs for uploads and downloads.
• aws___get_tasks: polls the status of long-running tasks started by aws___run_script.

Serverless capability. AWS also documents a read-only serverless capability, scoped to your own account, with helper functions diagnose, search_logs, get_live_config, get_recent_changes and get_trace_summary for working out why a Lambda function is failing across SQS, SNS, DynamoDB, API Gateway, EventBridge and Step Functions.

Governance features. The server adds two IAM global condition context keys, aws:ViaAWSMCPService and aws:CalledViaAWSMCP, to every request so you can treat agent actions differently from direct API calls in IAM policies. CloudTrail logs all API calls, and the server publishes CloudWatch metrics.

Setup

The AWS MCP Server runs in eight Regions, each with its own endpoint, for example https://aws-mcp.us-east-1.api.aws/mcp (also us-west-2, ap-southeast-1, ap-southeast-2, ap-northeast-1, eu-central-1, eu-west-1 and eu-west-2). There are two ways to authenticate.

Option A: OAuth (simple). Attach the AWSMCPSignInOAuthAccessPolicy managed policy to your IAM role or user, then add the remote URL. For Claude Code the documented command is:

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

Claude Desktop, Cursor, Kiro IDE, Gemini CLI and Codex use the same URL with ?oauth=initialize appended. Sign-in happens in your browser through AWS Sign-in. Access tokens are valid for 1 hour and refresh automatically for up to 12 hours.

Option B: SigV4 (advanced). Install the AWS CLI (2.32.0 or later), run aws login, install uv, and connect through the MCP Proxy for AWS over stdio:

{
  "mcpServers": {
    "aws-mcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy-for-aws-cli@latest",
        "https://aws-mcp.us-east-1.api.aws/mcp",
        "--metadata", "AWS_REGION=us-west-2"
      ]
    }
  }
}

AWS's decision guide points to SigV4 when you need several AWS accounts in one session, a default Region, or read-only mode that hides write-capable tools from the agent. Without AWS_REGION, operations default to us-east-1.

Plugins and the CLI wizard. In Claude Code you can install everything at once with /plugin install aws-core@claude-plugins-official. With AWS CLI 2.35.0 or later, aws configure agent-toolkit detects installed agents such as Kiro, Cursor and Claude Code, installs default skills and configures the server connection.

If you already use the older AWS API MCP Server or AWS Knowledge MCP Server, AWS recommends removing them from your client config when you switch, to avoid tool conflicts.

Use cases

• Ask which AWS services fit a workload, such as a serverless API, using current docs and skills.
• Create resources like an S3 bucket with versioning and a lifecycle policy through authenticated API calls.
• Troubleshoot a failed CloudFormation deployment or a failing Lambda function with logs, config and recent changes.
• Check whether a service, feature or CloudFormation resource is available in a given Region.
• Run multi-step or cross-service checks in a sandboxed Python script with aws___run_script.

Limitations and things to watch

• Older preview labels. AWS announced general availability on May 6, 2026, after a preview that started at re:Invent 2025. The awslabs/mcp README still calls the server "in preview" and the MCP Registry entry still carries the version 1.0.0-preview.1, so treat the Agent Toolkit for AWS user guide as the current reference.
• Rate limits. Authenticated requests are limited to 10 per second per account per Region, and unauthenticated requests to 5 per second per source IP. Throttled requests get a 429 response. Neither quota is adjustable.
• Connection quotas. 100 concurrent connections per account per Region (not adjustable), 180 concurrent active sessions per account and 90 per user per Region (adjustable).
• Ephemeral storage. Session storage is kept for at most 8 hours.
• Unauthenticated access is limited. Without credentials you only get the read-only knowledge tools.
• OAuth limits. Multi-profile switching across accounts is not supported with OAuth, and organizations that restrict signin:AuthorizeOAuth2Access and signin:CreateOAuth2Token must use SigV4.
• Real AWS access. API tools act with your IAM permissions. AWS recommends scoping IAM roles down to the minimum the agent needs.

Pros and cons

Pros

• Official and run by AWS, with nothing to host.
• One endpoint covers docs, skills and authenticated API calls.
• IAM condition keys, CloudTrail logging and CloudWatch metrics give clear control and audit.
• Docs search works without any AWS credentials.
• Broad client support, plus plugins and a CLI wizard for quick setup.

Cons

• IAM setup is required before any API action.
• Fixed request rate limits per account and per IP.
• AWS routes multi-account sessions and read-only mode to the SigV4 proxy, which adds uv and the AWS CLI.

Who it is for

Developers and platform teams who build on AWS with an AI coding agent and want it to use current AWS docs and act on real accounts under IAM control. It suits teams that need an audit trail of agent actions and want to separate agent calls from human calls in IAM policies.

Pricing

As of October 2026, AWS states that you can use the Agent Toolkit for AWS, which includes the AWS MCP Server, at no additional charge. You pay only for the AWS resources your agent creates or uses, at standard AWS pricing.

Alternatives

• Azure MCP server for the equivalent official server on Microsoft Azure.
• Cloudflare MCP server if your stack runs on Cloudflare.
• Terraform MCP server if you manage AWS through infrastructure as code.

Sources

• What is the Agent Toolkit for AWS: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/what-is-agent-toolkit.html
• Setting up the AWS MCP Server: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/getting-started-aws-mcp-server.html
• Understanding the MCP Server tools: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/understanding-mcp-server-tools.html
• Capabilities: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/capabilities.html
• Quotas for AWS MCP Server: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/aws-mcp-limits.html
• Getting started: https://docs.aws.amazon.com/agent-toolkit/latest/userguide/quick-start.html
• AWS What's New, AWS MCP Server general availability: https://aws.amazon.com/about-aws/whats-new/2026/05/aws-mcp-server/
• awslabs/mcp README: https://github.com/awslabs/mcp
• Agent Toolkit for AWS repository: https://github.com/aws/agent-toolkit-for-aws