Source: microsoft/azure-skills · Maintainer: Microsoft · License: MIT · Checked: 28 October 2026
What is the AI Runway AKS setup skill?
The airunway-aks-setup skill is Microsoft's guided runbook for serving an open model on Azure Kubernetes Service (AKS) with AI Runway. It takes a coding agent from an existing cluster to a model that answers chat requests, in six fixed steps, and asks you before every install or deployment.
Its description reads: "Set up AI Runway on AKS" and says it "Covers cluster verification, controller install, GPU assessment, provider setup, and first deployment." Trigger phrases in the description include "deploy model to AKS", "KAITO setup on AKS", "vLLM on AKS" and "run LLM on AKS".
AI Runway itself is a separate open-source project. Its README describes a web UI and "a unified Kubernetes CRD (ModelDeployment) to deploy models across multiple inference providers", lists KAITO, NVIDIA Dynamo, KubeRay, LLM-D and direct vLLM as providers, and calls itself a Cloud Native Computing Foundation sandbox project. It also warns that "the APIs are not currently considered stable."
The skill ships in microsoft/azure-skills, the Azure Skills Plugin. The README lists it under "Build, deploy, and evolve" next to azure-prepare and azure-kubernetes. The SKILL.md frontmatter gives version 1.1.2 and author Microsoft.
As of 28 October 2026, its skills.sh page listed about 361,700 installs (snapshot, 28 October 2026).
What does it do?
The SKILL.md is short, about 4.5 KB. The detail lives in a references/steps/ folder with one file per step, which the agent loads only when it reaches that step. The skill uses no MCP tools; everything runs through kubectl, make and curl.
Five rules govern the run: go in order, report each check as healthy or missing, "Ask for user confirmation before any install or deployment action", skip steps that are already done, and honor skip-to-step N.
| Step | What the agent does |
|---|---|
| 1. Cluster verification | Checks that kubectl, make and curl exist, reads the current context, lists nodes and counts GPUs from NVIDIA device plugin labels |
| 2. Controller install | Looks for the modeldeployments.airunway.ai CRD; if missing, runs make controller-install and make controller-deploy and waits for the rollout |
| 3. GPU assessment | Matches each GPU model to a profile table (VRAM, bfloat16 support, attention backends) and works out the largest model that fits |
| 4. Provider setup | Recommends an inference provider, then runs make deploy in that provider's folder and checks it reports ready |
| 5. First deployment | Suggests a starter model, creates a Hugging Face token secret if the model is gated, applies a ModelDeployment and watches it reach Ready |
| 6. Summary | Recaps the setup, sends one chat completion to the endpoint and lists next steps |
Provider choice. The step 4 table defaults to KAITO, with llama.cpp for CPU-only clusters. It suggests NVIDIA Dynamo for "High-throughput serving with separate prefill/decode phases" and KubeRay for teams "Already using Ray for ML workloads". The skill's table covers those three; AI Runway's README also lists LLM-D and direct vLLM.
Model sizing. references/model-sizing.md gives rough weight sizes per parameter count at float16, int8 and int4, and says to add 20 to 30 percent for KV cache and activations. Its starter table ranges from a quantized Gemma 3 1B on CPU to Llama 3.1 70B across four A100 80 GB GPUs with tensor parallelism.
Old GPU warnings. For T4 and V100 nodes, the skill warns that they lack bfloat16 and tells the agent to add --dtype float16 to the serving arguments.
Secrets handled carefully. For gated Llama models, step 5 reads the Hugging Face token with a silent prompt, writes it to a temp file with chmod 600, creates the hf-token secret from that file and deletes the file. A powershell-notes.md file gives PowerShell versions of the Bash-specific commands in steps 1, 4, 5 and 6.
Rollback. references/troubleshooting.md maps common failures (CrashLoopBackOff, ImagePullBackOff, a 401 from Hugging Face) to fixes and lists undo commands for each step, from kubectl delete modeldeployment back to make controller-undeploy.
How do you install the AI Runway AKS setup skill?
The repo README documents several routes. Because this skill uses no MCP tools, the skill-only install should be enough on its own.
Skill only. The command on skills.sh copies just this folder:
npx skills add https://github.com/microsoft/azure-skills --skill airunway-aks-setup
Claude Code: the Azure Skills Plugin. This installs every Azure skill plus the Azure MCP Server. Run inside Claude Code:
/plugin install azure@claude-plugins-official
Or run /plugin and search for "azure". See Claude Code plugins.
GitHub Copilot CLI:
/plugin marketplace add microsoft/azure-skills
/plugin install azure@azure-skills
Codex, Cursor and Gemini CLI. For Codex, the README says to run codex plugin marketplace add microsoft/azure-skills, then install azure from /plugins. For Cursor, use the Cursor Marketplace or Settings, then Plugins. For Gemini CLI, run gemini extensions install https://github.com/microsoft/azure-skills. The README also documents apm install microsoft/azure-skills for Microsoft's APM tool. See Codex skills and Cursor skills.
What you need on your machine. The skill checks for kubectl, make and curl, and needs a kubeconfig context for your cluster (the skill suggests az aks get-credentials). Step 2 says the make targets "must be run from the AI Runway repository root". The skill has no clone step, so Kipt's reading is that you should clone ai-runway/airunway first and start the agent in that folder.
How do you use it?
Describe the goal in plain language. The skill is model-invoked, so the agent loads it when your request matches its trigger phrases.
Example prompts (suggestions):
Set up AI Runway on my AKS cluster and deploy a first model.Check whether my current AKS cluster is ready for GPU inference with AI Runway. Don't install anything yet.Run LLM inference on AKS with KAITO and vLLM. My nodes are A10s.Resume the AI Runway setup: skip-to-step 4.Deploy microsoft/Phi-3-mini-4k-instruct to AKS through AI Runway and give me the endpoint.
Expect a status report after each step, a yes or no question before each make or kubectl apply, and a final summary with the endpoint and a test request. Big models take a while: the skill says 70B models can need 20 to 40 minutes or more to download weights, and tells the agent to read pod logs rather than assume failure.
Tips and limits
From the repo, Microsoft Learn and skills.sh, not from testing:
- GPU nodes cost money. The skill opens with a cost warning and asks the agent to confirm you understand GPU charges before it provisions GPU resources. Check Azure's pricing for your VM size and region first.
- Check GPU quota. Microsoft's AKS AI toolchain operator page notes your subscription needs GPU VM quota in the cluster's region. The Azure quotas skill can check it.
- T4 and V100 with KAITO. The skill's starter table pairs a single T4 with KAITO and vLLM. AI Runway sets up upstream KAITO from the kaito-project Helm chart (its KAITO Makefile has a
setup-kaitotarget), not the AKS managed add-on. Even so, both KAITO's Azure docs and Microsoft's page for the managed AI toolchain operator add-on say GPUs older than Ampere aren't supported; Microsoft's page names "the T4, V100, M60, and K80" and says "KAITO requires NVIDIA GPUs with CUDA compute capability 8.0 or later." If your nodes are T4 or V100, check this before you follow that row. - AI Runway is young. Its README says the APIs aren't stable yet, so the
airunway.ai/v1alpha1manifests in the skill may change. - Day-2 incidents are out of scope. For problems with an existing GPU deployment, the skill points to an optional
aks-gpu-inferenceskill. That skill isn't part of the base azure plugin: it ships in the repo's separate, optional aks-skills plugin, and the skill tells the agent to ask before installing it. - Audits. On skills.sh, Gen Agent Trust Hub rated the skill Pass (Safe), Socket Pass and Snyk Pass (low risk), all on 9 October 2026.
The MIT License lets you copy and change the skill as long as you keep Microsoft's copyright and license notices.
Save it in your Kipt library
A copy you can edit makes sense if your team always uses one provider, one namespace or a fixed list of approved models. In Kipt, open Skills and paste the skill's GitHub folder link into the import box, or drop the folder or a .zip. Skills kept in your private GitHub repo install into Claude Code or Codex with one command:
npx skills add https://github.com/<you>/<repo>.git --all -g
Kipt is not affiliated with Microsoft or the AI Runway project.
Similar skills
- Azure quotas skill: checks GPU VM quota and requests increases before you add GPU nodes.
- Azure diagnostics skill: the baseline troubleshooting skill this one lists as related.
- Microsoft Foundry skill: deploys models on Microsoft Foundry instead of your own cluster.
- Hugging Face skill: teaches the hf CLI for finding and downloading the models you serve.
- Kubernetes manifest generator skill: writes the Deployments and Services around your model endpoint.