Security and AI choices

Know where your data goes—and what your agent is allowed to do.

We document what stays local, what may reach an AI provider, which tools are connected, and which actions always require your approval.

Discuss your requirements

The essential distinction

The agent does the work. The AI model helps it think.

OpenClaw, Hermes, or an evaluated NVIDIA NemoClaw-based stack can coordinate the workflow: watching for an approved event, gathering information, and preparing the next step. A selected hosted or local model helps interpret information and draft responses.

Because the agent foundation and AI model are separate choices, selected work can stay on your hardware while other tasks use a hosted model when it provides better results. We explain that data path before anything is connected.

Agent foundation

OpenClaw

An open-source assistant infrastructure option for scheduled work, connected tools, memory, and chat-based control.

Agent foundation

Hermes

An alternative agent option we can evaluate when its tool use, model support, and deployment characteristics fit the workflow.

Evaluated early-preview stack

NVIDIA NemoClaw

NemoClaw adds OpenShell sandboxing, policy controls, and routed inference around supported agents. It is not itself a model or a replacement for OpenClaw. Compatibility review is required, especially for Mac deployments.

Early-preview alpha · availability and platform support may change

AI on demand

ChatGPT or Claude

Helpful for drafting, research, and analysis when a person opens the app and asks for help.

Best for: Everyday individual or team assistance
Where data goes: Content is processed by the selected service
Tradeoff: Simple to use, but usually waits for a person to ask

Fully managed online

Hosted Agent

Custom workflows, memory, schedules, and a dashboard that stay available around the clock on a dedicated server.

Best for: 24/7 work without office hardware
Where data goes: Agent records stay on the server; approved requests may go to the chosen AI provider
Tradeoff: Fast and flexible, but still cloud-hosted

Runs from your office

Private Node

An in-office agent and storage layer that can use local AI, selected hosted AI, or both.

Best for: Greater local control, office files, or limited-connectivity work
Where data goes: Selected work can stay local; connected tools still receive the information they need
Tradeoff: More control, with hardware and backups that still need to be managed

Your model plan

Use the model that fits the task—not a permanent default chosen for every job.

We can evaluate models from OpenAI, Anthropic, Google, NVIDIA, xAI, Mistral, Meta, and other compatible providers, along with local and open-weight options. The recommendation is based on real output quality, privacy, latency, cost, tool support, and hardware.

Compare on your work

Models are tested against representative examples and an agreed quality bar.

Route intentionally

Routine classification, complex reasoning, sensitive processing, and fallback may use different options.

Keep an exit path

Where practical, we avoid architecture that makes changing a provider unnecessarily difficult.

Controls we design around

Always available.
Still under your control.

The right safeguards depend on the task. Reading a calendar is different from sending an email. Drafting an invoice is different from issuing a payment.

Separate agent accounts

The agent uses its own approved accounts instead of an employee’s personal login wherever possible.

Access only to what it needs

Reading, drafting, sending, deleting, publishing, and payments are treated as separate permissions.

Human approval

Important or irreversible actions stop for review unless a narrow, tested rule explicitly allows them.

Secure account connections

Credentials are protected and connected after delivery using limited access where the software supports it.

A record of what it did

Important actions can be logged, reviewed, and retained according to the policy you choose.

Backups and a shutdown plan

Updates, backups, monitoring, restart behavior, and an emergency stop are defined before launch.

Four questions to ask

Know where the work happens.

01Where does the agent run?
Provider workspace, dedicated VPS, office hardware, or a combination.
02Where is memory stored?
Local disk, managed database, provider feature, or connected business system.
03Which model sees the request?
A local model, OpenAI, Anthropic, another provider, or policy-based routing.
04Which tools receive information?
Email, CRM, calendar, chat, web services, and every other connected system.

Business data protections are stronger than many people assume—but details matter.

OpenAI and Anthropic state that business and API inputs and outputs are not used to train their models by default. Retention, consumer settings, feedback, and specific features can differ. We evaluate the exact product and configuration, not merely the vendor name.

Primary sources

OpenAI enterprise privacy OpenAI API data controls Anthropic training policy Anthropic retention controls NVIDIA NemoClaw overview NemoClaw platform support OWASP excessive agency guidance

Important

No system is invulnerable, and local installation alone does not establish regulatory compliance. HIPAA, PCI, GDPR, and other regulated deployments require use-case-specific review, eligible vendors and features, appropriate agreements, and operational policies. Compliance work is included only when stated in a signed scope.