If you unboxed an OAK camera and your first instinct was to point your coding agent at it, this release is for you, and we think you are going to love it.
Today we are releasing Luxonis Agent Skills: a focused set of skills that teach coding agents like Claude Code, Codex, and Cursor how to work with OAK cameras and DepthAI the way we do. They take you from a camera still in the box, to a verified working device, to a running proof of concept, faster than you thought possible. We believe this is the very beginning of a fundamental shift in how our users build on the platform, and we are excited to get it into your hands. It is a beta, and we are shipping it early on purpose. More on that at the end.
Why we built this
Engineers in robotics, automation, inspection, and spatial AI increasingly live inside a coding agent. But a general-purpose agent does not know our platform out of the box, so it guesses at APIs, mixes up DepthAI versions, and spends your time and tokens rediscovering what we already know cold. Agent Skills change that. Each skill primes your agent with the right approach, context, and guardrails for one specific job, so it behaves like someone who has built on OAK a hundred times. The platform's full capability is available the moment you plug in, through the agent you already use.
The skills
Five skills follow one simple path: get the camera working, shape your idea into a brief, build the demo. Troubleshooting sits alongside, ready whenever you need it.
luxonis-guide is the router. Not sure where to start? Run this and it points you to the right skill.
luxonis-device-setup brings a camera to a verified, working state. Every setup is a little different, and networking is where it shows most: USB, PoE, Ethernet, switches, VPNs, and link-local addressing can each change the path to a connected device. This skill lets your agent navigate both our tools and your specific setup to find the best route to a working camera, then saves a DEVICE.md so later sessions start from a known-good state.
luxonis-project-interview turns an idea into a buildable brief. The hardest part of working with any agent is getting what is in your head into a form it can act on. This skill does exactly that: it interviews you one question at a time in plain language, then writes a PROJECT_BRIEF.md that becomes the shared context every later step, and every future session, can build from.
luxonis-build-poc takes that brief and stands up the thinnest runnable slice that proves the core idea on a real device. It adapts the closest known-good Luxonis example rather than inventing a pipeline from memory, confirms it is targeting DepthAI v3, and verifies the result with a command you can re-run. Watching a working OAK demo come together in minutes lands even with the expert engineers here who built the platform.
luxonis-troubleshoot is the "something is wrong" entry point, for a broken pipeline, bad depth, low FPS, or a device that will not show up. It reproduces the problem, isolates device versus app, diffs your app against a known-good example, and changes one thing at a time until the symptom is gone. If it cannot fix things locally, it writes a clean SUPPORT_PACKET.md for our team.
How the skills stay grounded: luxonis-context
What makes an agent great on OAK is context, so the skills give it real, current sources instead of relying on memory. When a skill needs something real to build or diff against, it uses our curated oak-examples repository, cloned once into a shared folder at ~/.luxonis/agent-context/ and reused across your projects. And for any Luxonis-specific fact, the skills start from our documentation map at docs.luxonis.com/llms.txt and follow its Markdown pages rather than guessing. That llms.txt map is useful well beyond these skills: point any LLM or agent at it directly and you will get far better answers about our platform. We recommend it.
What this is, and what it is not
luxonis-build-poc exists to help you validate a use case fast, to answer "can an OAK do this, and roughly how" in an afternoon. It builds the thinnest slice that proves the idea, not a production system, and luxonis-troubleshoot can help you tune performance along the way. The goal is to let fast-moving R&D teams prove out a project in hours, then decide where to invest.
We are hopeful this unlocks a whole new way to build perception and robotics systems, dramatically lowering the barrier to entry and shrinking time to value. We are confident that with tooling like this, hardware teams can start to move as fast as software teams. This is just the first step, we have a lot in the pipeline, and we would love your input along the way.
A quick note on where this fits next to Luxonis Chat: for quick questions and simple lookups, chat.luxonis.com is still the right tool. For anything where an agent is actually building, connecting, or debugging, these skills are a big step up. It is the same split you already make between Claude and Claude Code: one for quick answers, the other for doing the work.
Installing the skills
No server to stand up, no heavy install. The skills are deliberately small, to keep setup trivial and token usage low, and they work across whatever harness you prefer. You bring the agent you already like, the skills bring the Luxonis knowledge.
The repository lives at github.com/luxonis/skills.
Claude Code, run these in your session:
/plugin marketplace add luxonis/skills
/plugin install luxonis@luxonis
/reload-plugins
Cursor, install from the Cursor Marketplace, or add manually via Settings, Rules, Add Rule, Remote Rule (Github) with:
luxonis/skills
Any other agent (Codex and others) via the skills CLI:
npx skills@latest add luxonis/skills
Using the skills
Every skill is manual-only, invoked by name, and never triggers on its own:
/luxonis-guide
/luxonis-device-setup
/luxonis-project-interview
/luxonis-build-poc
/luxonis-troubleshoot
Claude Code plugin installs may namespace commands by plugin name, for example /luxonis:luxonis-device-setup.
A typical first session: /luxonis-device-setup to verify your camera, /luxonis-project-interview to turn your idea into a brief, /luxonis-build-poc to stand up the demo, and /luxonis-troubleshoot if you get stuck. For hardware faults or anything the skills cannot resolve locally, reach us at support@luxonis.com.
This is a beta, tell us everything
We are releasing this early because the fastest way to make it great is to watch real engineers use it. We want feedback on all of it: ease of install and use, the quality of what it built, how many tokens it burned, where it got stuck, what capabilities you wish it had, and which integrations would help.
This is the very beginning of a much larger effort, and we believe it points to where building on OAK is headed. Install it, point it at a camera, build something, and tell us how it went. We cannot wait to see what you make.
Please give us your thoughts below in the comments section or, if you have private information, directly to our PM: max.mclaughlin@luxonis.com
Get Started: https://github.com/luxonis/skills
