Nur AI

NurGlobal AI chat and coding assistance for VS Code. Independent software, with no OpenAI affiliation.

Install the supplied VSIX using Extensions: Install from VSIX. Open the Nur AI sidebar and choose Sign in. Approve the displayed application and scopes in your browser. Credentials use VS Code SecretStorage.

Select code and run Nur AI: Ask About Selected Code, or explicitly choose a file. Repositories are never uploaded automatically. Ignored files and common secret paths are excluded from agent tools. Agent changes require review/apply approval; checks require exact-command approval. Workspace Trust is required for agent execution. Stop cancels the current request or check. Saved agent sessions can be resumed without automatic replay of uncertain mutations.

The panel has three modes (Chat, Coding agent, Builder) and four tabs (Chat, Work, Console, Preview). Alongside them:

• Nur AI: Edit Selection Inline (Ctrl+I / Cmd+I on a selection) proposes a replacement, shows it in the native diff editor, and applies it only after you confirm.
• Nur AI: Toggle Inline Completion (Tab) enables gray-text completions as you type (nurAI.inlineCompletion.enabled, off by default). Completions need a signed-in account and a trusted workspace.
• Nur AI: Add Folder as Context, Add Documentation as Context and Search Repository as Context attach labelled context from a bounded workspace index, so a question can reference real paths and line numbers. Context is additive and is only sent when you send the message.

This extension uses the connected NurGlobal server for local-model inference. It is not a general VS Code extension host or a promise of guaranteed code correctness. No marketplace publication is performed.

Controlled browser and live workflow (1.5.0)

Use Nur AI: Open Controlled Browser (or Open browser in editor) for a full editor-sized browser beside code. The Preview tab also contains a controlled Chromium browser with separate Reference and Project preview tabs. Navigate, go back/forward, reload or stop, select a 1280×800 desktop or 390×844 mobile viewport, capture a masked screenshot, and add observed page context to the active task. Click the page capture to interact; Inspect DOM provides selectors and private typing controls. Captures update after controls/actions; this is not a continuous video stream.

The agent uses the same browser through browser_action for navigation, DOM/style inspection, scrolling, clicking, typing and screenshot interpretation. compare uses matching viewports and sends reference/preview screenshots to an advertised vision model. Missing/failed vision is explicitly reported as unverified. Page content is evidence, not permission or trusted instructions. Reference interactions ask for approval; credentials, MFA and CAPTCHA use User login / MFA handoff. Finish handoff, then send continue. Native security-key prompts, protected media, anti-automation challenges and sites that prohibit headless browsers may require an external browser and remain unverified in the controlled session.

Browser cookies/local storage/IndexedDB are kept in VS Code SecretStorage, scoped to the workspace and tab; no plaintext authentication profile is written into the project. Input values are omitted from DOM context and masked in captures. Use Clear sessions / log out to delete saved browser sessions. SessionStorage is not persisted. Do not place credentials in URLs or task messages.

The extension runs as a workspace extension. In Remote SSH, Chromium, commands and previews run on the SSH host. Its view travels through the VS Code webview message channel; no public debugging/browser port is opened. Open forwarded URL uses vscode.env.asExternalUri for an external browser. An external browser has a separate login session and is not controlled by the agent.

Playwright Core is included in the VSIX. An installed Chromium/Chrome is detected where supported, or set nurAI.browser.executablePath to its path on the workspace host. Nur AI: Install Browser Runtime installs only Chromium using Playwright's installer. OS libraries may require administrator installation; failed startup reports the dependency error. Browser binaries are not bundled into the VSIX.

Start project preview detects static HTML, Vite, Next.js or Laravel. The exact loopback command requires approval; no dependency installation is implicit. Other frameworks need their own approved startup command and a loopback URL. Stop preview process or Stop cancels managed processes; extension shutdown cleans them up. Previews are restarted explicitly after an extension-host restart.

File paths in activity are clickable. Follow edits opens files with preserveFocus and can be disabled. Changed-file counts derive from successful writes. State, objective, observations, reference URLs and execution journal survive reload via workspace state, an atomic private extension-storage checkpoint, and backend checkpoints. A completed response is not automatically a completed project. waiting_for_user identifies login/MFA handoff separately from tool approval.

Build from the existing repository:

npm ci --prefix source/packages/client
npm ci --prefix source/clients/extension
npm run build --prefix source/clients/extension
npm run package --prefix source/clients/extension
code --install-extension source/clients/extension/nur-ai-1.5.3.vsix --force

Run Developer: Reload Window. For Remote SSH install into the host that owns the workspace. The backend needs the matching checkpoint state update accepting waiting_for_user; no database migration is added by 1.5.0.

Image attachments (1.5.1): paste clipboard images, drop images onto the composer, or use + Image to select a file on the client computer. These transfer bytes, not client paths, including in Remote SSH. To select an image on the workspace host instead, run Nur AI: Attach Image from Workspace Host. PNG, JPEG and WebP are supported up to 20 MB and 25 megapixels, six tray items. The tray shows upload progress, thumbnails for client and workspace-host uploads, and retry/remove controls. Pending bytes are stored privately in workspace-scoped extension storage; successful uploads retain owned asset IDs. Removing a tray item does not delete already recorded task history or the backend asset. Starting a new task clears the tray. Images require a model advertising vision; a failed or unavailable interpretation is reported explicitly. Saved observations and asset IDs survive reload and are reused on continuation without another upload.

Version 1.5.2 adds workspace-host thumbnails, checks provider-reported size before reading selected host images, and enforces tray capacity across concurrent uploads.

Browser repair candidate 1.5.3: Windows acceptance is still pending. Browser discovery now prefers the matching Playwright-managed Chromium and checks per-user Chrome, Program Files / Program Files (x86) Chrome and Edge. An explicit invalid executable reports an error; it is not silently ignored. Install Browser Runtime installs Chromium on the workspace host. Restart is not needed after changing the executable setting; the next launch reads it.

Go, Start project preview and Open Live Preview use the same controlled session as the agent. Sidebar/editor share selected tabs and replay state when shown. Captures refresh once per second while idle and after actions; this is a remote rendered view, not a full native browser. No address-bar extensions, downloads, popups, video-rate streaming or OS dialogs are provided. Click and scroll the image; use Inspect DOM then Type privately / Enter for form fields. Sensitive input values stay masked. Public-site compatibility is not guaranteed; startup, navigation and capture errors are displayed in the browser view.

For Windows/XAMPP preview, set nurAI.preview.phpExecutable to C:\xampp\php\php.exe if PHP is not on PATH. Standard drive-root XAMPP is also detected. Startup output/errors appear below the page. No application files are changed by preview startup itself. Test this release in a separate fixture first.

If browser/preview fails, run Nur AI: Collect Browser Diagnostics in the affected workspace window after reproducing it. Save the generated JSON tab and share it. It records extension/host/runtime versions, executable selection, launch/capture probe, timestamped UI render acknowledgements and action errors; no credentials, input values, images or environment dump. A successful capture is not proof the image decoded in a hidden view: inspect the render timestamps.

Personal agents (1.6.0)

Choose Personal agent above the composer: Coder, Builder, Planner or Reviewer. Planner and Reviewer receive only file listing/search/reading and analysis-plan tools. Runtime dispatch also rejects edits, shell commands, preview startup and browser actions for these profiles. Start a New task to change agents. The profile is saved with the local task binding so resuming keeps the same access.

Use Manage to create, edit or delete up to 20 personal agents. Give each a name, instructions and either reviewed coding or read-only access. These preferences are private editor user state, not a public agent marketplace. Instructions are sent to your connected Nur AI server when used; do not put credentials in them. Changing instructions never expands workspace permissions.

Project rules opens .nur/rules.md in your workspace. Saved rules (up to 12KB) are loaded for each coding-agent turn using the same ignored-file and workspace-containment checks as file tools. Rules express conventions, not authorization. Unsaved changes are not sent. An ignored rules file is skipped. Existing browser, checkpoints, image attachments and approval workflows remain available to Coder/Builder. Personal agents run one at a time; they are not parallel worker processes.