NVIDIA Agent

NVIDIA Agent is a VS Code AI coding agent powered by NVIDIA NIM and OpenAI-compatible chat models. It is built for developers who want more than a chatbot: ask it to inspect a workspace, edit files, create projects, improve UI, run commands, review changes, and keep working until it has a useful result.

Homepage | Marketplace | Sample prompts

Independent extension. Not affiliated with NVIDIA Corporation. Requires your own NVIDIA NIM-compatible API key.

Why NVIDIA Agent

NVIDIA Agent brings an agentic workflow directly into VS Code:

• It understands natural language tasks.
• It reads relevant workspace context.
• It plans the next action.
• It edits files through controlled tools.
• It shows progress while working.
• It summarizes what changed.
• It lets you review or undo generated edits.

The goal is simple: describe the outcome you want, and let the agent do the repetitive coding work inside your editor.

Key Features

Agentic Coding

• Autonomous agent loop: plan, act, observe, fix, and finish.
• Real workspace actions: list files, read files, write files, patch files, create folders, run commands, start background commands, and open browser previews.
• Mutation guard: for edit/create/fix/refactor/style requests, the agent keeps working until a real file change happens or a real blocker is found.
• Planning timeout protection: prevents long silent model stalls.
• Deterministic fallback paths for common tasks such as CSS polish when the model hesitates.
• Final summaries with changed files, line counts, verification status, and recommended next steps.

Multi-Agent Workflow

Enable the Multi-agent team mode to run specialist passes before producing the final answer:

• Architect: plans architecture and integration.
• Coder: writes implementation-focused changes.
• Reviewer: checks bugs, safety, regressions, and tests.

NVIDIA Agent can automatically fall back to single-agent mode when the token budget is too tight.

Realtime Chat Experience

• Realtime streaming responses when the selected endpoint supports streaming.
• Working animation while the agent runs.
• Natural progress narration while it works, such as what it is reading, why it continues, and which file it edited.
• Detailed timeline events for explored files, edited files, commands, and reasoning checkpoints.
• Stop button for cancelling a running request.

Code Generation And Autocomplete-Style Help

NVIDIA Agent is designed for fast coding assistance from the chat panel and editor actions:

• Generate new files or whole project scaffolds from a prompt.
• Add, rewrite, or refactor code in existing files.
• Explain selected code from the editor context menu.
• Fix selected code from the editor context menu.
• Generate tests from the current editor.
• Use chat as an autocomplete-style coding partner when you want a larger change than inline completion can safely handle.

Native inline ghost-text autocomplete is not the main surface yet; the current strength is agentic chat, editor commands, and workspace edits with review/rollback.

Review And Rollback

Every generated file edit can be snapshot-backed:

• Review edited files after the agent finishes.
• See per-file diff stats such as +143 -3.
• Open changed files directly from the review picker.
• Undo the last generated edit with rollback.
• Use permission modes to decide when edits are applied.

Model Support

• NVIDIA NIM OpenAI-compatible chat endpoint support.
• Default endpoint:

https://integrate.api.nvidia.com/v1/chat/completions

• Built-in model catalog.
• Live model refresh from /v1/models.
• NVIDIA Build catalog supplement.
• Manual model ID input for compatible providers.
• Auto Strongest Free model routing mode.
• Model picker in the chat composer.

API Key Management

• Secure API key storage using VS Code SecretStorage.
• Save one key or a pool of NVIDIA NIM API keys.
• Optional automatic key rotation for quota, rate-limit, auth, timeout, or model availability errors.
• API status indicator in the header.

Token Saver

NVIDIA Agent includes local token controls to reduce cost and waiting time:

• Eco: short history, smaller workspace context, lower token use.
• Balanced: practical default for normal coding work.
• Power: larger context and more agent work for difficult tasks.
• Configurable context budget.
• Configurable max history turns.
• Configurable max workspace file count.
• Prompt size estimate before requests.
• Auto-disable multi-agent mode when the request is over budget.

Permissions

Choose how much autonomy the agent has:

• Ask before edits: confirm generated file edits and risky actions.
• Workspace trusted: automatically apply generated workspace edits.
• Administrator trusted mode: trusted automation mode for repositories you control.

Use trusted modes only in workspaces where you are comfortable letting an AI coding agent modify files.

Attachments And Context

• Attach workspace files as context.
• Upload text files, PDFs, images, audio, and other file metadata.
• Paste images from the clipboard into the chat.
• Include lightweight workspace context automatically.
• Clear conversation context when you want to reduce hidden history.

Chat History

• New chat button.
• Recent task list.
• Reopen previous conversations.
• Persisted chat sessions in VS Code global state.
• Clear conversation when needed.

Usage Tracking

• Local request count.
• Prompt token count when returned by the API.
• Completion token count when returned by the API.
• Total token count.
• Clear usage button.

Multilingual UI

Choose the extension language:

• Auto
• English
• Vietnamese
• Chinese
• Spanish

The assistant response language follows the selected language unless the user asks for another language.

What You Can Ask

Create a React todo app, install dependencies, run the build, start the dev server, and open the browser preview.

Read this workspace, identify the main app structure, improve the CSS, and summarize what changed.

Fix the build error, inspect the failing output, patch the source, and run the build again.

Review this extension architecture and identify reliability, security, and token-cost risks.

Add dark mode, persist the setting locally, and keep the edits rollback-safe.

Generate tests for the selected file.

Quick Start

1. Install NVIDIA Agent from the VS Code Marketplace.
2. Open the NVIDIA Agent icon in the Activity Bar.
3. Open the API Key tab.
4. Paste your NVIDIA NIM API key and save.
5. Choose a model, or keep Auto Strongest Free.
6. Ask the agent to build, fix, review, explain, test, or improve your project.

Settings Overview

NVIDIA Agent exposes its main controls in the extension UI and VS Code settings:

• Endpoint
• Model
• Max tokens
• Temperature
• Streaming responses
• Workspace context
• Thinking mode
• Multi-agent team
• Requests per minute
• Permission mode
• Enabled specialist agents
• Token Saver mode
• Context token budget
• History turns
• Workspace file count
• Auto single-agent fallback
• Auto API key rotation
• Extension language

Strengths

• Good for project-level requests, not only single snippets.
• Safer than blind generation because file edits are reviewable and rollbackable.
• Cost-aware through Token Saver and context trimming.
• Flexible model routing through NVIDIA NIM compatible endpoints.
• Friendly for multilingual teams.
• Useful for repeated workflows: build, fix, test, review, refactor, and UI polish.
• Designed to feel like an agent working inside VS Code, with progress narration and visible work history.

Current Notes

• Native inline ghost-text autocomplete is not the primary feature yet.
• Agent quality depends on the selected model and API endpoint reliability.
• Some models may be slower or may not support streaming.
• For high-risk repositories, keep permission mode on Ask before edits.

Privacy And Keys

• API keys are stored in VS Code SecretStorage.
• Keys are used only for your configured NVIDIA NIM/OpenAI-compatible endpoint.
• Chat and workspace context are sent to your configured model endpoint when you ask the agent to work.
• You control permission mode before file writes and automation.

License

NVIDIA Agent is licensed under the NVIDIA Agent Personal Use License.

Personal, individual, non-commercial use is allowed. Commercial use, redistribution, resale, sublicensing, rebranding, marketplace republishing, paid hosted services, and commercial derivatives are prohibited without prior written permission from the copyright holder.

See LICENSE for the full terms.