Jetson Edge AI Development: Setting Up a VS Code + Claude Code Workstation

Background: The Unique Challenges of Edge AI Development

NVIDIA Jetson is a series of embedded system modules designed for edge AI computing. Each module integrates GPU processing power, memory, and I/O onto a single board, typically consuming between 10–60W. This makes Jetson ideal for deployment in robots, drones, industrial cameras, and other scenarios requiring local real-time inference. Unlike cloud-based GPU servers, Jetson's resource constraints and proprietary software stack (JetPack/CUDA/TensorRT) make configuring the development toolchain a challenge in itself.

Developing on edge AI devices often means dealing with limited compute, cumbersome remote connections, and an unfamiliar toolchain. JetsonHacks has put together a highly practical tutorial on transforming VS Code into an AI-powered programming workstation for NVIDIA Jetson. By combining Claude Code with NVIDIA Skills (device skills) introduced in JetPack 7.2, developers can achieve a complete development loop: edit on the host machine, run on Jetson, and invoke AI coding agents from either side.

This article breaks down the setup approach for this workflow and highlights several practical tips for working with coding agents.

Environment Setup: Making VS Code the Foundation for Jetson Development

The tutorial is built around a Jetson AGX Orin with 64GB of memory, connected to a keyboard, mouse, display, and a local network with internet access. The Jetson AGX Orin is the flagship of the series, featuring a 12-core ARM Cortex-A78AE CPU and a 2048-core Ampere architecture GPU, delivering up to 275 TOPS of AI performance. The 64GB version uses unified memory (LPDDR5X) shared between CPU and GPU, enabling it to run fairly large language models locally. An Orin Nano on the same network is introduced later to demonstrate headless remote development.

The first step is installing VS Code. The author opts to download the ARM64 Debian package and install it with dpkg from the download directory — currently the most popular code development environment on Jetson.

On the choice of coding agent: Claude Code is a command-line AI coding agent from Anthropic, built on the Claude model. It can read and write files, execute terminal commands, search codebases, and complete multi-step programming tasks. Unlike inline completion tools such as GitHub Copilot, a coding agent has autonomous planning and execution capabilities — it can take a high-level goal, break it into steps, call tools, and iterate until the task is complete. The author admits he has no strong preference among coding agents, but a user poll on the official Jetson AGX platform showed respondents preferred Claude Code by a 2:1 margin, making it the natural focus for this tutorial. He also shares a practical strategy: Claude Code requires a paid subscription, while OpenAI Codex tends to be more token-efficient with a more generous quota. His approach is to subscribe to the lowest paid tier of both and "let them complement each other" rather than upgrading to a higher tier.

Token Budget Management: Using the Research and Projects Features Wisely

Tokens are the basic unit language models use to process text — roughly equivalent to 0.75 English words or 0.5 Chinese characters. Every interaction with a model, including both your input and the model's output, consumes tokens. With a paid subscription, you can choose your preferred model and reasoning depth. "Reasoning depth" refers to the internal chain-of-thought steps a model takes before delivering a final answer. These intermediate steps consume tokens but can significantly improve answer quality on complex problems.

The author cautions that learning to allocate your token budget wisely takes time. Spending is subject to both 5-hour and weekly caps, and poor allocation can leave you "out of fuel" at a critical point in a project. Smart strategies include reducing reasoning depth during research, using the Projects feature to build up context and avoid re-entering it each session, and choosing lighter model variants for tasks that don't require maximum capability.

The Research feature works like a smarter web search. Instead of returning a list of links, you can ask a direct question and the model will synthesize results from across forums and multiple sites into a direct answer.

But here's a critically important warning: blindly trusting an answer just because it "looks good" is the most common trap. The risk of this agentic workflow is that each error the agent makes gets amplified by subsequent steps, creating a cascade of hallucinations. The right approach is to treat the AI's answer as a starting point for finding the truth — go back and verify sources, keep asking follow-up questions, and make sure you genuinely understand. As the author puts it: "Remember, an LLM gives you an answer, not the answer."

The Projects feature is often underestimated. It lets you group multiple conversations, carry context across an entire project rather than a single session, and continuously add background information. Claude, ChatGPT, and Grok all offer similar functionality. This feature is especially effective during the research phase and when writing specification documents.

Installing the Claude Code Extension and NVIDIA Skills

Anthropic provides an official Claude Code extension for VS Code, installable through the standard extension marketplace. After installation, click the Claude Code tab, follow the prompts to connect your subscription account, complete the web-based authorization, and switch back to VS Code to start using it.

NVIDIA Skills, introduced in JetPack 7.2, is an AI tool capability extension framework for the Jetson platform. At its core, it's a set of predefined prompt templates and tool-calling interfaces that allow AI coding agents to understand and operate Jetson-specific system-level features. These skills cover typical scenarios such as: diagnosing GPU/CPU/memory usage via jtop or tegrastats, managing unified memory allocation, optimizing TensorRT engine serialization and model serving configurations, and managing Docker-based NGC AI container ecosystems. Compared to general-purpose coding agents, the value of device skills lies in encapsulating platform-specific knowledge — such as Jetson's memory architecture differences and JetPack version compatibility — into directly callable capabilities, reducing the need for developers to repeatedly consult hardware documentation.

For simple repositories, you can clone and install directly. A more efficient approach is to first have the AI research and explain a tool, then have it generate a prompt, and pass that to an agent like Claude Code to perform the installation.

The author describes "having one AI generate a prompt to pass to another AI for execution" as many people's first "aha" moment in their AI journey. Once installed, users can interact with Jetson in natural language — whether building a project or troubleshooting an issue — in a far more intuitive way.

Core Philosophy: The Agent Works for You, Not the Other Way Around

If there's one principle from this entire tutorial worth remembering, it's this: the agent works for you — you don't work for it.

The author repeatedly emphasizes that using a coding agent requires human guidance and oversight. Left to its own devices, an agent can "produce garbage at an order of magnitude faster than you can clean it up." This isn't an exaggeration — agentic automation by nature streamlines human workflows, and without clear direction, technical debt accumulates at the same accelerated pace. Investing time upfront in specification writing and project planning yields enormous returns.

This points to a reality that's easy to overlook: AI coding agents excel at rapid output, but the quality of that output depends entirely on the user's ability to guide them. Automation without direction simply accumulates technical debt faster.

Headless Development: SSH Connections and VS Code Remote Extension

A headless configuration means the device runs without a monitor, keyboard, or mouse, receiving instructions solely over the network. This is the standard deployment mode for embedded and edge computing scenarios, since peripherals add power consumption, size, and cost. About half of Jetson developers work this way — device connected to power and network only. In this setup, SSH (Secure Shell) is the most fundamental interaction method, establishing an encrypted tunnel on port 22 that supports terminal command execution and file transfer.

Before connecting, you need to know the Jetson's IP address or hostname on the local network. If unknown, use nmap to scan the subnet (nmap -sn 192.168.1.0/24 quickly lists all active devices on the LAN along with their IPs and hostnames; install via apt if not already present). The author suggests that troubleshooting network issues is a great use case for AI — models have extensive training data on these topics and typically give reliable guidance. Once you can successfully ping and SSH into the device, you're ready to work on the remote Jetson.

A handy tip: when switching between the host machine and the target device, it's extremely easy to lose track of which terminal is which. Use different colors for different terminal windows to keep things straight.

Configuring Remote Development in VS Code

After installing Microsoft's Remote-SSH extension, you can work with code on a remote Jetson directly from your host machine. The extension silently installs the VS Code Server (ARM64 version) on the remote machine, rendering the editor UI locally while all file operations, terminal processes, and debug sessions run on the remote Jetson. This gives developers a complete local IDE experience while code runs directly in Jetson's hardware environment — eliminating the "works on my machine" problem.

Use the command palette to run Remote-SSH: Connect to Host, enter the remote Jetson's IP address, save the configuration, and a new window connected to the remote environment opens.

It's easy to confuse local and remote VS Code windows — there's a badge in the lower-left corner to distinguish them, and the author recommends setting a different color theme for the remote environment as well. A terminal opened in the remote window points directly to the remote machine, so you can install Jetson device skills on the Orin Nano directly from there.

Running Claude Code in the Terminal

Beyond the VS Code extension, Claude Code also supports a pure terminal mode. SSH into the remote device, launch Claude, complete the initial configuration and account binding, and you have full functionality in a CLI environment. You can invoke Jetson device skills using slash commands or ask questions in natural language. For developers comfortable with the command line, this approach is often more efficient — especially on a resource-constrained Orin Nano, where CLI mode has significantly lower overhead than a graphical IDE.

Closing Thoughts: The More Powerful the Tool, the More Critical Your Judgment

The author closes with a well-measured reminder: information about these tools comes from sources with widely varying levels of experience and credibility. Without hands-on experience, it's hard to truly make good use of expert advice — but at the same time, you shouldn't blindly follow people who only have a superficial understanding of the underlying mechanics.

The "VS Code + Claude Code + NVIDIA Skills" combination is fundamentally about lowering the barrier to entry for Jetson edge AI development. But what ultimately determines success is still the developer's command of these tools and their ability to critically verify outputs. The more powerful the tool, the more indispensable human judgment becomes.

Key Takeaways