Screenshots

Nyx AI on Windows 11.

Real captures of signed Nyx AI builds, not mock-ups. Nothing here is relabelled or recoloured: each shot names the build it was taken on, and the interface, controls and version stamp are exactly what that build showed. A shot may be newer than the release currently delivered by the Microsoft Store.

The app at rest

Chat mode on an empty session. The sidebar carries sessions, projects and scheduled tasks; the footer shows the selected model and the current approval mode. The edition badge, workspace name and app version sit in the title bar.

Captured on 1.5.486

Chat mode builds a document

A one-page PDF requested in chat. Nyx AI shows its reasoning, the write step as a labelled tool chip, and the finished file as a download card. Chat mode is deliberately document-only: no shell, no git, no workspace file tools, so nothing it does here can reach your projects.

Captured on 1.5.484

Nyx AI Lab reads the project

Lab mode opens a workspace. Here it reads index.html, summarises what the page contains, and groups the other files by type — including ones that are not code. Each tool call is shown as its own chip you can expand.

Captured on 1.5.484

A risky command asks Windows, not the page

The model proposed a push Nyx AI rates high-risk. High- or critical-risk model-authored shell commands require a native Windows confirmation naming the exact command, workspace, reason and requester, or are refused; everyday commands follow your approval mode instead. The AI cannot answer that dialog.

Captured on 1.5.484

Four isolation levels, and what each one enforces

Settings → Safety. Automatic, Standard, Sandbox + internet and Sandbox · no internet — each row states in the app's own words what it enforces, including where that level stops. The Security page sets out the boundaries and the residual risks in full.

Captured on 1.5.484

You pick the model

The model picker showing cloud entries. Models with a known licence caveat are marked “⚠ licence terms” — a prompt to check with your provider, not an acceptance flow in the app. Local models served from your own machine appear in the same picker.

Captured on 1.5.484

Scheduled tasks, with the cloud caveat in view

Interval, hourly, daily, weekdays, weekly, monthly or a cron expression. The dialog warns that an unattended task's prompt and the context it reads may be sent to its model host, and defaults to the safest approval setting. The dialog was restyled after this capture.

Captured on 1.5.484

From a project note to a rendered document

Lab mode on a workspace. One request produced three labelled tool calls you can expand — it read notes.md, wrote the PDF, then opened it — and the finished document is rendered in the pane on the right, served from a loopback address inside the app rather than handed to an external reader. The model supplied the wording; the section structure came from the file it was asked to read.

Captured on 1.5.493

These show a subset of the app. The feature and getting-started pages describe what is not pictured here, including language servers, MCP servers and the two audit logs.