Only your text prompt leaves your machine. Scripts, output, credentials, and session data never do. Two modes, one guarantee — you decide what crosses the boundary.
What you see in the chat is the full conversation, including every script output. What the AI receives is a separate, automatically managed context — usually shorter, trimmed to fit the model's window. Only your prompts and the outputs you actively send via Send to AI enter it.
Solid arrow: your prompt — crosses automatically.
Dashed arrow: a specific script output — crosses only when you click Send to AI on that result.
Reply: the generated script, returned to the chat.
The client verifies the proxy. The proxy verifies the device. Both directions, every call — not just at login. The verification logic lives in the Native C++ Core and protects only the integrity of the connection.
Enterprise SSL inspection sees your prompt, the generated script, and a short-lived session token bound to this specific device. None of those lets an attacker impersonate the proxy or reproduce the backend. A captured token without the matching device is inert.
The Secure Proxy strips your identity before forwarding the request. Your email, subscription, and company are not part of what Google sees — only an anonymous device hash and your prompt.
Prompts, scripts, chat history, and script outputs are not persisted on the proxy or anywhere else. The proxy holds only what it needs to operate: your email, subscription state, and the device hash.
In Local AI mode, the client sends prompts directly to the OpenAI-compatible endpoint you configured — LM Studio, Ollama, Azure OpenAI, or any other. The ShellPilot proxy is not in the inference path — prompts, scripts, and outputs never leave your network for AI processing. Script output is opt-in here too: it crosses only when you click Send to AI — which is why you can switch between Cloud and Local mid-session without leaking anything.
You choose the context length in the settings — 16K, 24K, 32K, 64K, or 128K tokens, matched to your model and hardware. The AI context is trimmed to exactly the limit you selected, separate from the visible chat.
While a Local AI session is running, the client refreshes its licence token hourly when online. If the proxy is briefly unreachable, the session continues for up to 24 hours since the last successful refresh. The proxy never receives prompts, scripts, or output in this mode.
Every fresh launch of ShellPilot contacts the Azure proxy once to verify your subscription before the main window opens, regardless of mode. After that, in Local AI mode the inference path stays inside your network.
Designed for environments with strict data-sovereignty requirements: banks, government, healthcare, regulated industries. AI inference, prompt content, and script output all stay on your network. Also your business-continuity path: a cloud outage does not stop your infrastructure management.
Any OpenAI-compatible endpoint works — your model choice, your hardware, your hosting. ShellPilot does not impose a model or a vendor for the Local AI path. Local models have inherent limitations (smaller context window); optimized for standard scripting and automation tasks. Available on Pro and Pro Plus.
AMSI, App Control, script-block logging, execution policy — the security page answers reviewer questions precisely.