The personal_assistant profile strips AWS API access down to Bedrock inference and nothing else. The agent can call foundation models, but it cannot read or write any other AWS service. This makes it the right choice when you want an AI assistant for general tasks — writing, research, coding help, planning — without any risk of the agent interacting with your AWS account’s resources.
What personal_assistant grants
The personal_assistant profile uses a single inline policy with four permission groups:
| Permission group | Actions |
|---|
| Bedrock inference | bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:GetUseCaseForModelAccess |
| Bedrock discovery | bedrock:ListFoundationModels, bedrock:GetFoundationModel, bedrock:ListInferenceProfiles |
| SSM connectivity | SSM Session Manager actions (required to connect to the instance) |
| Identity | sts:GetCallerIdentity |
When to use personal_assistant
Use personal_assistant for general-purpose tasks that don’t require AWS knowledge:
- Writing and editing — drafting documents, emails, reports, and summaries
- Research — synthesizing information, comparing options, answering questions
- Coding help — reviewing code, explaining concepts, writing scripts (without deploying anything)
- Planning — breaking down projects, generating checklists, thinking through architectures
- Daily tasks — anything you’d use a capable AI assistant for
Default instance size
Personal_assistant defaults to a t4g.medium (2 vCPU, 4 GiB RAM, arm64 Graviton).
Inference-only workloads don’t need a large instance. The t4g.medium handles them well and keeps costs low (~$25/month for the instance).
Security services
All security services are off by default with personal_assistant:
| Service | Default |
|---|
| Security Hub | Off |
| GuardDuty | Off |
| Inspector | Off |
| Access Analyzer | Off |
| Config recorder | Off |
Required for NemoClaw
personal_assistant is the only compatible profile for NemoClaw, the sandboxed agent pack.
NemoClaw runs OpenClaw inside an NVIDIA OpenShell sandbox that applies Landlock filesystem restrictions, seccomp syscall filtering, and network namespace isolation. The sandbox blocks all outbound AWS API calls from inside the container — so any profile that grants AWS access would be meaningless at best and confusing at worst.
With personal_assistant, the agent routes Bedrock inference through the bedrockify proxy running on the host (outside the sandbox), so it still gets model access without any direct AWS API surface inside the sandbox.
Do not use builder or account_assistant with NemoClaw. The sandbox will block the AWS API calls those profiles are designed to enable, and the deploy will not work as expected.
Deploy with personal_assistant
Claude Code
NemoClaw (sandboxed)
OpenClaw
Hermes
curl -sfL install.lowkey.run | bash -s -- -y --pack claude-code --profile personal_assistant
curl -sfL install.lowkey.run | bash -s -- -y --pack nemoclaw --profile personal_assistant
curl -sfL install.lowkey.run | bash -s -- -y --pack openclaw --profile personal_assistant
curl -sfL install.lowkey.run | bash -s -- -y --pack hermes --profile personal_assistant
You can deploy personal_assistant in your main AWS account — not just a sandbox — since the agent has no access to your AWS resources. It can only invoke Bedrock models and connect via SSM.
