Run Wingman skills

Skills are packaged Axelera workflows: one prompt in, an end-to-end result out. Watch a skill build and run a complete pipeline on your card, then fly it yourself from Claude Code or Agent Dev.

One prompt to wingman-launch: the skill picks a model, builds the pipeline, runs it on your card, and opens the result viewer.

What a skill is

A skill is a packaged set of instructions your AI CLI loads for a specific kind of task — launching a pipeline, deploying a model, benchmarking, debugging. Wingman ships 19 Axelera-aware skills. They know the Voyager SDK, the model zoo, and Axelera hardware, so the agent stops guessing and follows the proven path. Skills run wherever your CLI runs: in a Wingman-launched Claude Code terminal, in Agent Dev, or in a plain terminal outside the app.

Install and run skills

  1. Get the skills

    Two routes, same skills. The desktop app installs them for you on first run, into your home directory: ~/.claude/skills/ for Claude Code and $CODEX_HOME/skills/ (default ~/.codex/skills/) for Codex CLI. Or install the Wingman plugin for your CLI — see the plugins page for the one-time install. Either way the location is per-user, so the skills are there in every session and every project; nothing is copied into your own project folders. In Claude Code they appear as slash commands like /axelera-wingman:wingman-launch, Codex picks them up automatically, and in Agent Dev inside the desktop app they are already available to the agent.

  2. Run your first skill

    Start Claude Code from Wingman, then ask for an end-to-end result in plain language — build and run a person detection demo on my Axelera card — or invoke the skill directly. wingman-launch is the one-shot launcher: it selects the model, builds the pipeline, validates it on hardware, and packages a RUN_HERE.sh with logs and a browser result viewer. When the card is in a machine you reach over ssh, the package runs there as a device target; see Run on a device you cannot install Wingman on.

    Claude Code

    /axelera-wingman:wingman-launch build a person detection demo from my USB camera

    Plain prompt

    Build and run a person detection demo on my Axelera card, then show me the results in the browser.

  3. Pick the right skill for the task

    Skills are scoped, and the agent routes between them — but knowing the map helps you ask better. Use wingman-launch for anything that should end in a running pipeline. Use wingman-add-model when a model needs new integration metadata or authoring, and wingman-deploy to compile or quantize an existing YAML model, standalone recipe, or AxScript card. Use wingman-debug when you bring an existing error or a board that stopped responding. Use wingman-bench to measure FPS and latency on a pipeline that already works.

  4. Let skills and MCP tools work together

    Skills handle local execution; the hosted MCP tools handle knowledge and record-keeping. A typical run uses both — the skill searches the Voyager knowledge base through rag_search_knowledge_base, executes locally, and writes what it produced into a Wingman project with create_document. You get this combination automatically when you launch the CLI from Wingman.

Safe model inspection has a narrow boundary

How model skills choose a surface

Model surfaceOwning front doorRouting rule
Established YAML model zoodeploy.py and inference.pyStart from the nearest released YAML integration for supported ONNX, PyTorch, timm, or Ultralytics inputs.
Standalone recipe catalogReleased recipe or axzooFollow the selected recipe’s source, backend, and maturity contract. The selected v1.8.0 release labels this surface Alpha.
AxScriptpython -m model_axscriptUse compatible weights or a checkpoint plus a source implementation and authored graph or kernels. A bare ONNX file is not an AxScript input.

Let evidence set the capability claim

Which hardware a skill validates on

The 16 Wingman skills

SkillUse it for
wingman-launchOne-shot end-to-end pipelines and demos: build, run on the card, validate, package with RUN_HERE.sh and a result viewer, or run the package on a device target over ssh.
wingman-deployCompile, quantize, and calibrate an existing YAML model, standalone recipe, or AxScript card for the Axelera APU.
wingman-runRun an already-selected model or pipeline against a video, image, camera, stream, or dataset, or execute an exact published precompiled LLM prompt with axllm.
wingman-benchMeasure FPS, throughput, and latency; compare model variants on working pipelines.
wingman-debugFix existing errors, crashes, accuracy issues, broken installs, or an undetected Axelera board.
wingman-new-pipelineDesign a YAML pipeline configuration when you want YAML-only design, not a full run.
wingman-new-appWrite a standalone Python app around a validated pipeline with the InferenceStream API.
wingman-buildCompile native Voyager SDK components: GStreamer operators, trackers, C/C++ examples.
wingman-testRun pytest suites for Voyager SDK components.
wingman-add-modelInspect and route a custom artifact, then add the YAML, recipe, source mapping, AxScript graph, kernel, decoder, or evaluator the chosen surface requires.
wingman-list-modelsFind model zoo models by task, framework, or dataset.
wingman-voyager-helpExplanation-only answers about the Voyager SDK, hardware, and APIs — no code changes.
wingman-agentBroad orientation and multi-step planning that spans several skills.
wingman-init-wingmanSet up the Wingman MCP integration for a new Voyager SDK workspace.
wingman-changelogGenerate a changelog fragment for the current PR in the Wingman development workflow.
wingman-docsKeep a documentation page in sync with the feature it describes when the UI changes.
wingman-githubSearch and read code, issues, pull requests, releases, and files on GitHub, with a URL cited for everything used.
wingman-huggingfaceSearch models, datasets, Spaces, and papers on the Hugging Face Hub, read model cards and file lists, and download weights for a port.
wingman-screenTake a screenshot, record the screen or one window as a video, and drive a GUI window on your own desktop. Typing and clicking start only after you hand over the machine in that session. The capture scripts install with the desktop app.

Let the agent route