SparkCode

Local AI coding assistant powered by your NVIDIA DGX Spark.

SparkCode replaces OpenAI Codex with Nemotron Super 120B running on your own hardware via vLLM — no cloud, no rate limits, no telemetry. Everything stays on your network.

Features

Chat sidebar — persistent conversation history, streaming responses, code highlighting
Agent mode — give it a task, it reads files, writes code, runs commands, and shows diffs before touching anything
Inline completions — ghost-text suggestions as you type (Tab to accept)
Selection commands — right-click any code to explain, refactor, fix, or comment it
Git tools — agent can check status, diff, log, and commit with your approval
VS Code diagnostics — agent sees TypeScript errors and ESLint warnings
Prompt library — quick-access templates for common tasks
Token usage — see how many tokens each response used

Requirements

An NVIDIA DGX Spark running vLLM with Nemotron Super 120B
Tailscale connecting your devices to the DGX Spark
The DGX Spark must be reachable at the configured endpoint (default: http://spark-2441:8000)

The extension is fully functional on any device in your Tailscale network. On devices outside the network, the status bar shows a red ✗ and all features gracefully fail with an error.

Getting Started

Install the extension
Join your Tailscale network
The ⚡ Spark status bar item will turn teal when connected
Press Ctrl+Shift+S (or click the ⚡ icon in the Activity Bar) to open the chat

Usage

Chat

Ctrl+Shift+S or click the ⚡ icon in the Activity Bar.

Ask questions about your code. The active file is automatically included as context on the first message. Previous sessions are saved and accessible via the history button.

Agent

Switch to Agent mode in the sidebar, or press Ctrl+Shift+A.

Describe a multi-step task in plain English. The agent:

Plans the approach
Reads relevant files
Proposes changes — you approve or reject each write before it happens
Runs commands (with your approval)
Summarizes what it did

Selection Commands

Highlight any code → right-click:

Explain — plain English explanation with edge cases
Refactor — cleaned up version with change notes
Fix / Debug — bug list + fixed code
Add Comments — inline comments added

Inline Completions

Start typing — ghost text appears after a short delay. Tab to accept, Escape to dismiss.

Can be toggled via sparkcode.inlineCompletionEnabled.

Settings

Setting	Default	Description
`sparkcode.endpoint`	`http://spark-2441:8000`	vLLM server URL
`sparkcode.model`	`auto`	Model ID (`auto` uses first available)
`sparkcode.maxTokens`	`4096`	Max tokens per response
`sparkcode.temperature`	`0.2`	Sampling temperature
`sparkcode.agentRequireApproval`	`writes-only`	`always`, `writes-only`, or `never`
`sparkcode.agentMaxIterations`	`30`	Max agent loop iterations
`sparkcode.inlineCompletionEnabled`	`true`	Toggle ghost-text completions
`sparkcode.inlineCompletionDelay`	`600`	Milliseconds before triggering completion
`sparkcode.promptLibrary`	`[]`	Custom quick-prompt templates
`sparkcode.gitToolsEnabled`	`true`	Enable git tools in agent
`sparkcode.systemPrompt`	(see settings)	System prompt for all requests
`sparkcode.shellPath`	(auto)	Shell for `run_command` tool

Keyboard Shortcuts

Shortcut	Action
`Ctrl+Shift+S`	Open SparkCode sidebar
`Ctrl+Shift+A`	Run Agent Task

Agent Tools

The agent has access to:

Tool	Description
`read_file`	Read any file in the workspace
`write_file`	Write a file (shows diff, requires approval)
`list_directory`	List directory contents
`search_files`	Regex search across the workspace
`run_command`	Run shell commands (requires approval)
`get_diagnostics`	Read VS Code error/warning diagnostics
`get_workspace_structure`	Recursive file tree
`git_status`	Current git status
`git_diff`	Show unstaged or staged changes
`git_log`	Recent commit history
`git_commit`	Stage all + commit (requires approval)

Privacy

All requests go directly from VS Code to your vLLM server over Tailscale. No data is sent to any third-party service. No telemetry, no analytics, no cloud.

Publisher

Built by AMP Titans.