← Back to tutorials

AI-assisted development on NVIDIA Jetson

A New Way to
Build Edge AI.

Put coding agents alongside real Jetson hardware to inspect, prototype, benchmark, and optimize faster.

Less repetitive work. More time to build what matters.

Watch the walkthrough Try the prompts ↓

Why AI-assisted development

Turn device insight into results, faster.

A coding agent working with real Jetson hardware The coding agent harness runs on the Jetson alongside project files and device tools. It can read a connected camera and sensors, and issue approved commands to actuators through device software. Measurements feed the next experiment. RUNNING ON JETSON Coding agent harness Project files + terminal + device tools Inspect → Run → Measure Camera Sensors Actuators
Agent tools run on Jetson. AI models can run locally or in the cloud. You set permissions and safety limits.

A coding agent can do more than suggest code. Give it controlled access to your Jetson project and it can inspect the real system, carry out approved experiments, and help you interpret the results.

  1. Inspect the real device.

    Discover connected hardware, available memory, software, and logs.

  2. Run real experiments.

    Capture camera data, read sensors, and test approved actions through device tools.

  3. Build on measured results.

    Compare model performance, refine your prototype, and verify on the target hardware.

The development impact

Shorten the path from idea to deployment.

Spend less time preparing data, building tools, and repeating setup. Put more time into testing ideas and improving your product.

Compress the repetitive work. Preserve human judgment.

Conceptual comparison

Scroll horizontally to compare the full development cycle →

Conventional development

More manual preparation and repeated work

AI-assisted development

Less repetitive work; more focus on your product

More time for your product.
Earlier delivery.

Time protected for product development and quality. Agents can assist throughout; you guide the design, implementation, and release.

Illustrative workflow, not measured results or guaranteed time savings.

Developer time

Less repetitive setup, scripting, data preparation, and experiment orchestration. More attention for design and difficult problems.

Time to market

Test assumptions earlier with working prototypes, reusable tools, and faster model evaluation cycles.

Total cost of ownership

Engineering effort is one part of TCO. Evaluate it alongside hardware, agent and compute costs, validation, deployment, and ongoing support.

Choose your coding agent

Your Jetson. Your coding agent.

The workflow is agent-agnostic. Start with the recorded Codex example; additional assistant walkthroughs are planned.

Walkthrough 01 · Setup · Steps 1–7

From a USB cable to a ready-to-build Jetson.

Watch a Mac connect to a Jetson Orin Nano Developer Kit, prepare Codex, install Jetson Device Skills, and finish GPU container setup.

Before the first prompt, install the ChatGPT desktop app on your host and sign in with Codex access. The recording uses macOS; UI and connection details may differ on other hosts.

  1. Begin in a Mac-side task: connect, authenticate, and configure SSH.
  2. Install Codex CLI on Jetson and register its remote project.
  3. Inspect the device, add Skills, and prepare Docker with approval.
Follow the setup prompts ↓
Watch on YouTube ↗

Directly on Jetson

Install and sign in to Codex CLI on Jetson, then run it from your project directory in a local terminal or SSH session. The agent's commands and files are on the device.

From your desktop app

Use a host-side task for initial setup over SSH. Then add the Jetson in Settings > Connections in the ChatGPT desktop app and choose a remote project folder. Start the application task there.

Keep the host and Jetson awake and connected during this walkthrough. For unattended workloads, configure and test persistent services explicitly; an open SSH session alone is not a persistence guarantee.

Codex CLI setup ↗ SSH remote projects ↗ Jetson Agent Skills →

Companion walkthrough planned

Claude Code

A step-by-step Jetson example is not yet available here. The shared approach stays the same: provide device context, review the plan, run bounded experiments, and validate the result. Setup and available integrations depend on the assistant.

Explore Jetson Agent Skills

Companion walkthrough planned

Cursor

A step-by-step Jetson example is not yet available here. The shared approach stays the same: provide device context, review the plan, run bounded experiments, and validate the result. Setup and available integrations depend on the assistant.

Explore Jetson Agent Skills

Companion walkthrough planned

Devin

A step-by-step Jetson example is not yet available here. The shared approach stays the same: provide device context, review the plan, run bounded experiments, and validate the result. Setup and available integrations depend on the assistant.

Explore Jetson Agent Skills

Companion walkthrough planned

Local coding LLMs

A step-by-step Jetson example is not yet available here. The shared approach stays the same: provide device context, review the plan, run bounded experiments, and validate the result. Setup and available integrations depend on the assistant.

Explore Jetson Agent Skills

Prompt examples

Start with an outcome. Let the agent help with the steps.

These public prompts preserve the intent of the two recorded walkthroughs. Adapt the user name, connection, and model to your device. They are starting points, not unattended scripts.

Before you start: device, access, and approval checklist

Device and connection

  • Complete Jetson's first-run setup and reboot. Keep a local display and keyboard available.
  • Use a data-capable cable and the USB device-mode port, or connect by Ethernet/Wi-Fi. Seeing the L4T-README drive does not establish SSH connectivity.
  • Keep DHCP automatic. On macOS, allow the USB accessory and check the app's Local Network permission if Terminal can connect but the app cannot.
  • The recorded R39.2.1 setup used a USB-network workaround before recording. If USB networking fails, resolve it first or use Ethernet; the prompts are not a universal USB repair procedure.

You stay in control

  • Use a least-privilege account. Verify host identity, preserve existing SSH keys, and investigate changed-key warnings.
  • Enter passwords and sign-in codes locally. Never commit credentials or include them in shared prompts.
  • Read proposed commands and review downloads, service changes, and reboots before approval.
  • Prompt wording is not a security boundary. Configure the agent's actual permissions and verify what it changed.
Need to set up your developer kit first?

Part 01 · Steps 1–7

Connect and prepare your Jetson

Same host-side setup task
Step 1Connect to your Jetson

Recorded on macOS over USB. Replace “Mac” and the login name for your setup. Confirm the intended device before accepting an SSH host key; investigate changed-key warnings.

Adapt, review, then send 
Connect this Mac to my USB-connected Jetson and set up key-based SSH.

1. Guide my first login as jetson in this app's terminal. If using screen, guide login and cleanup.
2. Use jon-xxxx for the SSH alias and HostKeyAlias, using the last four hex digits of the built-in Ethernet NIC's permanent MAC. Use a dedicated SSH key.
3. Confirm access works. Keep DHCP automatic; Wi-Fi is next.

Step 2Give Jetson its own internet connection

Skip this step if the Jetson already has independent internet access. Do not paste Wi-Fi passwords into chat or project files.

Adapt, review, then send 
Connect the Jetson to its own Wi-Fi, keeping USB SSH working.

Guide me to enter the SSID and password locally, not in this chat. Enable automatic reconnection and verify internet access without this Mac.

Keep credentials only in the protected Wi-Fi profile.

Step 3Prepare Codex on the Jetson

Follow the official installation and sign-in flow. Enter passwords and authorization codes only in the appropriate local terminal or sign-in page.

Adapt, review, then send 
Prepare this Jetson for a Codex remote project.

Use our SSH connection to install Codex CLI if needed, guide sign-in, and prepare ~/jetson-ai-demo.

Connection registration is next.

Step 4Add the remote project

In the ChatGPT desktop app, SSH hosts are managed in Settings > Connections. Register the project, but keep this setup conversation on your Mac through step 7.

Adapt, review, then send 
Add this Jetson's SSH connection and ~/jetson-ai-demo as a remote project in this app. Guide any required clicks.

Finish when it appears under Projects. Keep this task on the Mac.

Step 5Establish a read-only baseline

Understand the actual device before changing it. Missing host AI libraries are not necessarily blockers when the application uses containers.

Adapt, review, then send 
Create a baseline for this Jetson's container-based VLM demo.

Summarize hardware, software versions, memory, storage, and GPU container readiness with the desktop running. Focus on actual deployment blockers, not missing host AI libraries.

Read-only; no downloads or changes.

Step 6Add Jetson Device Skills

Install on the Jetson, not the Mac. The Jetson Agent Skills guide covers installation paths and the full catalog. Start a fresh Jetson-side task for the application walkthrough.

Adapt, review, then send 
Install NVIDIA Jetson Device Skills from the official repository for the jetson user's Codex on the Jetson, not this Mac.

Confirm they are installed in the correct location.

Step 7Finish GPU container setup

Review downloads and system changes before approving. If Docker group membership changes, verify access from a fresh remote session before starting the application.

Adapt, review, then send 
Read the Jetson Device Skills over SSH and finish this Jetson's Docker setup for GPU-accelerated AI demos.

Propose only missing setup, then apply after approval. No full JetPack SDK; swap only if needed. Use a small compatible container to test GPU access.

Keep the desktop and USB SSH working.

Part 02 · Steps 8–10

Build a live VLM demo

New task in the Jetson remote project

Already have a prepared Jetson? Start here. Connect the webcam and display to Jetson, verify GPU container access in the new session, and keep these three prompts in the same task.

Step 8Plan a live VLM application

Start a new task inside the Jetson remote project. Attach the USB webcam and display to Jetson first. Check the selected runtime against your Jetson model and software release.

Adapt, review, then send 
Plan a live VLM demo for this Jetson using an existing app and prebuilt containers.

- USB webcam and display attached to the Jetson; browser capture is fine
- Inference on the Jetson; localhost access only

Find an app through Jetson AI Lab. Use https://www.jetson-ai-lab.com/models/gemma4-e2b/ for a lightweight, vision-capable setup that fits alongside the desktop.

Show the plan and download sizes before deploying.

Step 9Deploy and check real camera inference

Approve the plan first. Use an image that includes the required upstream features. Five minutes is a smoke test, not proof of production reliability.

Adapt, review, then send 
Deploy the approved VLM demo.

Show working captions from the Jetson's USB webcam in its browser. Check GPU acceleration, latency, and memory stability for five minutes.

Use the existing WebUI; no custom UI or Mac preview.

Step 10Make it a standalone kiosk

Use a dedicated demo account. Review auto-login and startup services, keep rollback instructions, and shut down safely before cycling power. Verify fresh captions, not just an open browser.

Adapt, review, then send 
Make this Jetson boot directly into the fullscreen VLM demo, with camera capture and inference running automatically, without this Mac or internet.

Reuse upstream kiosk support. Ask before enabling desktop auto-login or rebooting.

Guide a reboot and safe power-off/on test, confirm fresh captions without clicks, and leave rollback instructions.

Built with AI assistance

From a request to a standalone camera demo.

A worked example, not a blank-sheet rebuild: use Codex to assemble and configure existing software, then validate the result on a Jetson Orin Nano Developer Kit.

Watch on YouTube ↗

Walkthrough 02 · Application · Steps 8–10

Live VLM WebUI kiosk

A USB camera supplies live images to a local vision-language model. The Jetson displays fresh descriptions in its own browser, then starts the demo automatically after boot.

Reused building blocks
Live VLM WebUI, a Gemma4 E2B inference runtime, prebuilt containers, and upstream kiosk support.
Agent-assisted work
Select a compatible setup, configure services, connect the camera, check resource use, and prepare automatic startup.
Human validation
Review the plan, change the camera scene, check fresh captions, and verify reboot and offline startup. A demo is not production certification.
Live VLM WebUI guide → Gemma4 E2B setup → Upstream kiosk guide ↗

01

USB camera

Real scene

02

Local VLM

Inference on Jetson

03

Live WebUI

Fresh captions

04

Standalone kiosk

No laptop after setup

Bring your AI into the physical world.

Connect cameras and sensors. Build, test, and refine on real Jetson hardware.

Try the first prompt ↑