Today we’re releasing the Luxonis Agent Toolkit, the fastest way to start building with OAK.
Give the coding agent you already use the Luxonis context, tools, and workflows it needs to work with OAK. Then describe what you want to build. Whether you use Claude Code, Codex, or Cursor, your agent can help set up an OAK workspace, connect a device, find the right example or model, write an application, and work with the output from a real camera.
This is the easiest way to agentically code on the Luxonis platform, and the quickest path from an OAK on your desk to an application you can keep building.
We build OAK to make useful physical automation possible in more places. Cameras and AI can help machines see the work around them: goods moving through a warehouse, a product on a line, a robot navigating a space, or the small signals that tell an operator when something needs attention. But turning that potential into a working application has traditionally required a rare mix of computer vision, embedded, and software expertise.
Coding agents change what is possible. The Luxonis Agent Toolkit puts that change to work on the open, capable platform behind OAK.
A better way to start building with OAK
Useful physical automation starts with a system that can see, understand, and respond in the real environment. That is what OAK provides. OAK 4 brings synchronized cameras, depth perception, edge AI compute, and the software needed to build and deploy applications into one device. It can run applications standalone, without a separate host computer, or support a connected development workflow when that is the right fit.
That removes a large part of the hardware and systems-integration burden. It also gives builders room to create the application they actually need: from an off-the-shelf model and an existing example to a custom pipeline, custom model, or a standalone OAK application. DepthAI gives them the software foundation to work across that range.
But an integrated platform only feels simple when it is simple to build with. The remaining work is where teams often lose time: learning the stack, finding the right examples, connecting the device, and teaching a general-purpose coding agent enough about OAK to be useful. Without that context, an agent may find an old snippet, miss a relevant example, or write code without seeing what the camera is actually producing.
The Luxonis Agent Toolkit closes that gap. It gives your agent a working understanding of the platform: the practical workflows needed to build with OAK, plus the Luxonis MCP server for current documentation, examples, and model information.
The result is a much more direct path from a physical problem to a capable OAK application. You keep the coding environment you prefer; your agent gets the context it needs to help you build on the full Luxonis platform.
Start with the application you want to build
The Toolkit has one primary entry point: tell your agent what you want to make.
You might ask for a people counter for a doorway, a barcode reader for a production line, a depth-enabled robot behavior, or an object-detection application that publishes results to your existing system. The agent can turn that request into a project, identify what it needs to learn or verify, and bring in the right OAK-specific capability along the way.
That can include:
Preparing an OAK development workspace and getting a device ready
Finding current Luxonis documentation, known-good examples, and compatible models
Building and running a DepthAI application
Capturing a recording of the real scene so work can continue without the camera attached
Inspecting live frames, depth, detections, and pipeline output instead of guessing from logs
Troubleshooting an existing application or integrating a custom model
The agent does not force you through a fixed checklist. A quick first-run request and a more specific application request need different help. The Toolkit routes the work based on the task, while keeping the project context, notes, and plans with the application as it develops.
What’s in the Toolkit
The Toolkit includes one entry skill and seven specialists. You can ask for what you need in plain language. Your agent can select the relevant skill, or you can invoke one directly when you want a focused job.
luxonis is the starting point. It answers OAK questions, helps choose a device, checks that the workspace is ready, and routes work to a specialist.
luxonis-workspace prepares a folder for OAK development, including the tools, project environment, and context files that make later sessions easier.
luxonis-app builds or changes an OAK application, from a first capability to a fuller project.
luxonis-device-setup gets OAK hardware working for development and records the setup for the next session.
luxonis-record captures or replays a holistic recording of the real scene.
luxonis-inspect looks at what a running pipeline is actually producing with oakctl inspect.
luxonis-troubleshoot diagnoses an existing application that is failing or producing the wrong result.
luxonis-model converts and integrates a custom model that is not already ready in the Model Zoo.
The full plugin also bundles the Luxonis MCP server. That is what gives the agent access to current Luxonis documentation, examples, platform context, and model information while it works.
Built around the work that matters after “hello world”
The first version of Luxonis Agent Skills made it easier to get an OAK connected and adapt a relevant example. That remains a great way to learn the platform.
The Agent Toolkit takes the next step. It is built for the work of turning an idea into a real OAK application: writing the pipeline, running it against a scene, looking at the results, and iterating from evidence.
Two parts of V2 make a meaningful difference here.
The agent can work with the scene, not just the source code. With recording and replay, it can capture the camera inputs from a real environment and continue developing against that recording. With pipeline inspection, it can look at what the device is producing: frames, depth, detections, and other pipeline outputs. That gives the agent something far more useful than a terminal log when it needs to understand what happened.
The agent can use current Luxonis knowledge as it works. The bundled MCP server gives supported coding agents access to current platform documentation, examples, and model information. It reduces the scavenger hunt that normally happens when an agent tries to learn an unfamiliar hardware and software stack from the open web.
Together, these capabilities make OAK development feel much closer to what it should be: describe the application, get a useful starting implementation, run it, see the result, and keep improving it.
Built for the tools you already use
The Luxonis Agent Toolkit is deliberately agent-agnostic. The full plugin is available for Claude Code and Codex, Cursor can load the same standardized plugin layout locally, and the skills can be installed separately in other compatible coding agents.
The full plugin experience for Claude Code and Codex includes the Luxonis MCP server. This gives the agent direct access to the most relevant Luxonis sources while it is helping you build. The skills themselves are open source, so you can see how they work, adapt them for your own environment, and contribute improvements.
This matters because the best way to build a robotics or computer vision application is rarely to abandon the tools and workflow your team already uses. OAK should fit into that workflow. The Toolkit helps make that true from the first session.
From first build to production
The Toolkit makes an agent substantially more capable at working with OAK, and you can get surprisingly far with it. It can help turn an idea into a project, get the camera and application running, and give you a real base to improve.
As an application moves toward production, it deserves the same engineering judgment as any system that will run in the real world: a clear definition of success, an understanding of the environment, thoughtful model and architecture choices, and validation of performance. The Toolkit helps teams spend more of their time on those decisions and less of it fighting their way to a starting point. The skills are not intended to train models, build labeled datasets, invent SLAM systems, deliver complete ROS solutions, or certify anything for production. It builds and proves OAK applications, and can work alongside your other agents that handle those parts of the stack.
This is where we think the opportunity is biggest. More builders with a fast, open path to creating capable camera applications means more chances to put AI to work where it matters: alongside people, machines, and real operations.
Install the Luxonis Agent Toolkit
For the complete experience, install the plugin. It includes all eight skills and the Luxonis MCP server.
In Claude Code CLI, run these commands one at a time:
/plugin marketplace add luxonis/skills
/plugin install luxonis@luxonis
/reload-plugins
In Codex CLI, run:
codex plugin marketplace add luxonis/skills
codex plugin add luxonis@luxonis
Then start a new Codex session. Cursor setup and skills-only installation options are covered in the documentation.
For full installation instructions, see the Luxonis Agent Skills documentation. The source is available on GitHub.
Where this fits
For quick questions in the browser, Luxonis Chat remains the fastest path. The Agent Toolkit is your Luxonis assistant for local work: building, testing, debugging, and deeper questions inside the environment where your code lives. If you are writing DepthAI code, the Toolkit with your coding agent will beat copy-pasting from a chat window every time.
Still a beta, and moving fast
The Toolkit is in beta. The skills and the project files they create may change between releases, and you should review the generated plan, code, and evidence before relying on the result.
We want feedback on all of it: install experience, output quality, token usage, where it got stuck, and what you wish it could do. Open an issue on the repository, reach us on X @luxonis, or contact our PM directly max.mclaughlin[@luxonis.com](mailto:support@luxonis.com).
Install it, point your agent at a camera, and build something real.
Links
