Evidence

The checker leaves receipts for every letter. The coding agent that built it left receipts too. This page is the text record, so it can be read without screenshots.

Agent connection

Claude Code (the coding agent) was connected to the AWS account on 2 October 2026 with aws login, which signs the CLI in with AWS console credentials, and then to the managed AWS MCP Server through the official mcp-proxy-for-aws. It acts as a dedicated IAM user, plainly-agent, never as the account root. The account id is masked below.

$ aws login --region us-east-1 --profile plainly-agent
Updated profile plainly-agent to use arn:aws:iam::********1486:user/plainly-agent credentials.

# First call through the AWS MCP Server (tool aws___run_script -> sts:GetCallerIdentity), 2026-10-02 ~15:20 UTC
{"caller_arn": "arn:aws:iam::********1486:user/plainly-agent"}

Through the same MCP connection the agent then read back what it had shipped:

cloudformation:DescribeStacks  plainly       UPDATE_COMPLETE  (last update 2026-10-02T15:14:04Z)
apigatewayv2:GetApis           plainly-api   HTTP  https://3nf75pgrv4.execute-api.us-east-1.amazonaws.com
lambda:ListFunctions           plainly-api   python3.13  arm64  1024 MB
dynamodb:ListTables            plainly-data
cloudwatch:DescribeAlarms      plainly-api-5xx, plainly-lambda-errors, plainly-lambda-duration-p95,
                               plainly-bedrock-throttles  -> all OK

MCP configuration

The project-scoped entry in .mcp.json. It holds no secrets: it names the CLI profile the proxy signs requests with.

{
  "mcpServers": {
    "aws-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["mcp-proxy-for-aws-cli@latest", "https://aws-mcp.us-east-1.api.aws/mcp",
               "--metadata", "AWS_REGION=us-east-1"],
      "env": { "AWS_PROFILE": "plainly-agent", "AWS_REGION": "us-east-1" }
    }
  }
}

/mcp  ->  Reconnected to aws-mcp.

CloudTrail events

From CloudTrail event history (cloudtrail:LookupEvents filtered on user plainly-agent, itself run through MCP). Calls that went through the MCP server carry the MCP proxy and Claude Code in the user agent. The stack deploys ran from the agent's shell with the AWS CLI, so they show the CLI user agent. We list both rather than claim more than happened. Source IPs are left out.

Time (UTC)EventSourceUser agent
15:16:39DestroySessionaws-mcp.amazonaws.commcp-proxy-for-aws/1.7.0 claude-code/2.1.287
15:14:04ExecuteChangeSetcloudformationaws-cli/2.37.6 (deploy, second pass with the real URL)
15:11:32ExecuteChangeSetcloudformationaws-cli/2.37.6 (deploy, first pass)
15:07:07DeleteStackcloudformationaws-cli/2.37.6 (removing the stack CloudFront had blocked)
15:03:40ExecuteChangeSetcloudformationaws-cli/2.37.6 (attempt refused by CloudFront account verification)

Least-privilege policy

The agent's permissions (infra/agent-iam-policy.json, attached as plainly-agent-policy) cover only the services Plainly uses, only for plainly-* resources, only in us-east-1. Roles it creates must carry the plainly-boundary permissions boundary. An explicit deny applies whenever a call arrives through the MCP server (aws:ViaAWSMCPService is true): the agent cannot delete stacks, buckets, tables or IAM entities, stop or delete a CloudTrail trail, or create users and access keys.

We tested the deny on purpose. The agent asked, through MCP, to delete a stack:

cloudformation:DeleteStack  StackName=plainly-mcp-guardrail-test   (via AWS MCP Server)
AccessDenied: User: arn:aws:iam::********1486:user/plainly-agent is not authorized to perform:
cloudformation:DeleteStack on resource: arn:aws:cloudformation:us-east-1:********1486:stack/plainly-mcp-guardrail-test/*
with an explicit deny in an identity-based policy: arn:aws:iam::********1486:policy/plainly-agent-policy

The same user deleting a failed stack from the plain CLI at 15:07 was allowed. The deny is about the agent acting through MCP, which is the point. Creating IAM users and granting permissions was left to the account owner: Claude Code's own permission check stopped the agent from doing it, and the owner ran infra/setup-agent-user.sh instead.

Not done in time: a dedicated CloudTrail trail with MCP data events, and the deny-only lock policy described in infra/README-infra.md. Event history covers management events for 90 days.

Build log excerpt

From docs/agent-log.md. Times are 2 October 2026, UTC.

  • Before AWS: a research workflow on the rules, judges and field; 18 concepts scored; a parallel build with integration and adversarial review; scam rules tuned on a dev set and frozen before a single holdout run.
  • First AWS calls: the account is on the Free plan. Bedrock answered ValidationException: Operation not allowed and Textract SubscriptionRequiredException. The owner chose not to upgrade, so reading moved onto the device (Tesseract.js) and the rules engine runs alone. The AI path stays in the code, switched off.
  • Before 15:03 Two stack creates failed: the deploy role lacked apigateway:TagResource, which API Gateway needs when it creates a stage. Fixed in the policy file; the owner published version 2 at 15:02:50.
  • 15:03 to 15:07 The next create was refused by CloudFront: "Your account must be verified before you can add new CloudFront resources." The agent made CloudFront optional (UseCloudFront=false) and served the site from the same Lambda behind the HTTP API, with the same security headers and CSP.
  • 15:11 to 15:14 Deployed in two passes (the second rebuilds pages with the real URL). Live checks: every page returns 200, a scam SMS returns "Likely scam" with four flags.
  • ~15:20 AWS MCP Server connected; reads, the alarm check and the guardrail test above, all through MCP.