AWS AI

aws-ai

Use this skill to execute AWS CLI commands through the Docker image for the aws-ai helper (supported path). All agent-driven operations must use unattended mode (-e AWS_AI_UNATTENDED=true) to avoid blocking on MFA prompts.

When to use

• User asks to run AWS CLI commands via this project's helper.
• User wants the agent to operate AWS without installing AWS CLI locally.
• Session/MFA handling is needed (assume-role + temporary credentials).

Prerequisites

• Docker installed (docker --version).
• Environment file at ~/.aws-ai/.env with required variables:
  • AWS_AI_PROFILE
  • AWS_AI_ASSUME_ROLE_ARN
  • AWS_AI_MFA_SERIAL_ARN
  • AWS_REGION

Agent execution policy (always unattended)

First aws-ai run in the conversation

On the first time in the current chat session that the user asks you to run the helper (Docker flow), run docker pull for the image before any docker run, so the local image matches the registry (e.g. docker pull karibu/aws-ai:latest). Do this once per conversation session, not before every command.

The agent must always run the helper in unattended mode by passing -e AWS_AI_UNATTENDED=true in every docker run invocation. Do not use -i or -it: MFA is supplied via -e AWS_AI_MFA_TOKEN on retry, not from stdin, so a fully non-interactive docker run --rm (no TTY, no stdin attach) is enough.

Exit codes

Execution flow

1. Prime credentials (first time or after expiration)

Always use unattended mode. If MFA is needed, the helper exits with code 2.

docker run --rm \
  -v ~/.aws:/home/user/.aws \
  --env-file ~/.aws-ai/.env \
  -e AWS_AI_UNATTENDED=true \
  karibu/aws-ai:latest sts get-caller-identity

If exit code is 2:

1. Ask the user for their MFA code from their authenticator app. Do not ask them to paste it in chat.
2. Retry with the token provided via environment variable:

docker run --rm \
  -v ~/.aws:/home/user/.aws \
  --env-file ~/.aws-ai/.env \
  -e AWS_AI_UNATTENDED=true \
  -e AWS_AI_MFA_TOKEN=123456 \
  karibu/aws-ai:latest sts get-caller-identity

After successful priming, credentials are cached for the session duration.

2. Run AWS operations

Once primed (session cached), execute AWS commands in unattended mode:

docker run --rm \
  -v ~/.aws:/home/user/.aws \
  --env-file ~/.aws-ai/.env \
  -e AWS_AI_UNATTENDED=true \
  karibu/aws-ai:latest ec2 describe-instances

Note: Always include -e AWS_AI_UNATTENDED=true in every invocation. The helper will use cached credentials if valid; if expired, it will exit with code 2 and you must re-prime as described above.

3. Session refresh

Temporary STS sessions expire (duration configurable via AWS_AI_SESSION_DURATION_SECONDS, default 900 seconds). Track session age.

At 2 minutes before expiration (or immediately after detecting exit code 2 on any command):

1. Re-prime using the flow in step 1 (unattended mode, handle MFA code 2 if needed).
2. Only after successful re-priming, continue with AWS operations.

Error handling

MFA/session failures

If output indicates MFA/session issues (MultiFactorAuthentication failed, expired session token, invalid session) without naming a denied API action like service:Action:

• Re-prime with sts get-caller-identity using unattended mode.
• Retry the original command once.
• Do not confuse these with IAM permission errors.

Missing IAM permissions

If AWS returns AccessDenied, UnauthorizedOperation, not authorized to perform, or is not authorized to perform and the message names one or more API actions (e.g., ec2:DescribeInstances, iam:GetRole, s3:GetObject):

• Treat as insufficient policy on the assumed role.
• Tell the user exactly which Action(s) to allow in the IAM policy attached to that role.
• If the error includes a resource ARN, provide a minimal Statement example (Action + Resource / Condition as appropriate).
• If the message does not list the action, infer it from the CLI subcommand (e.g., aws ec2 describe-instances → ec2:DescribeInstances) and state any uncertainty.

Credentials persistence failures

If credential caching fails:

• Continue with the current command result.
• Warn that subsequent commands may require MFA re-priming.

Security guidelines

• Do not expose MFA codes, session tokens, or secret values in responses.
• Never ask users to paste MFA codes in chat. For the retry, they can export the code in their shell and pass -e AWS_AI_MFA_TOKEN (still without -i/-it).
• Summarize command outcomes clearly instead of dumping raw logs unless explicitly requested.
• AWS_AI_MFA_TOKEN is a single-use, short-lived value. Do not store it in .env or commit it.

Notes for human users (documented in README)

Human users can use interactive mode (-it) for manual MFA entry. This skill focuses on agent-driven unattended execution only. Refer to the repository README for human-interactive usage patterns.