Source: microsoft/azure-skills · Maintainer: Microsoft · License: MIT · Checked: 24 October 2026
What is the Microsoft Foundry skill?
The Microsoft Foundry skill gives Claude Code, GitHub Copilot and other coding agents step-by-step workflows for Microsoft Foundry: creating projects, deploying models, building and deploying hosted agents, running evaluations and fine-tuning models. It is for developers who build AI agents on Azure and want the coding agent to follow Microsoft's own procedures instead of guessing.
Microsoft Learn's page on the skill describes it as "a meta skill for Foundry work. It helps a coding agent choose the right Foundry workflow, load the matching sub-skill, inspect available Foundry MCP tools, and keep deployment and evaluation context consistent across turns."
It ships in microsoft/azure-skills, the Azure Skills Plugin. The README says the repo is synced automatically from microsoft/GitHub-Copilot-for-Azure, where contributions go. The SKILL.md frontmatter lists version 1.2.31 and author Microsoft.
As of 24 October 2026, its skills.sh page listed about 576,500 installs (snapshot, 24 October 2026).
What does it do?
The top-level SKILL.md is a router. It sets mandatory rules, then points to sub-skills for each job. The folder is large, about 220 files.
Mandatory rules, from the SKILL.md:
- Run
./scripts/check-and-setup-dependencies.sh(or.ps1on Windows) first and wait for it. The script exits ifazdis missing and installs themicrosoft.foundryazd extension if it isn't there. - Read the matching sub-skill document before calling any workflow-specific MCP tool.
- Call the Azure MCP
foundrytool first to discover the available Foundry MCP tools. - Read the bundled azd guidance before running any
azdcommand.
Sub-skills it routes to:
| Area | Sub-skills |
|---|---|
| Setup | project/create, resource/create, private-network (VNet isolation) |
| Models | models/deploy-model (preset, customized or capacity discovery), quota, rbac |
| Agents | create (quick start or advanced), create managed harness, create voice, deploy, invoke, routine, invocations-ws |
| Quality | observe (batch and continuous evals, prompt optimization), agent-optimizer, eval-datasets, insights, trace |
| Operations | troubleshoot, validate, cicd |
| Fine-tuning | finetuning (SFT, DPO and RFT with graders) |
It handles four agent types: prompt agents, hosted agents (your own container code), managed harness agents, and voice agents. Per Microsoft Learn on hosted agents, hosted agents "are your own code packaged as a container image" that run on Foundry Agent Service.
The skill also defines a .foundry/ folder inside each agent's source folder for local state such as agent-metadata.yaml, evaluation suites, datasets and results. It reads deployment values from azure.yaml and azd env get-values and avoids copying them into that metadata.
How do you install the Microsoft Foundry skill?
The Learn page and the repo README document these routes.
Claude Code: the Azure Skills Plugin. Run inside Claude Code:
/plugin install azure@claude-plugins-official
You can also run /plugin and search for "azure". The plugin bundles the Azure skills, including this one, with Azure MCP Server configuration. Update it with /plugin update azure@claude-plugins-official. Learn says to restart Claude Code if the Azure or Foundry tools don't appear. See Claude Code plugins and Claude Code MCP for how these pieces load.
GitHub Copilot CLI. From the README:
/plugin marketplace add microsoft/azure-skills
/plugin install azure@azure-skills
Codex. The README says to run codex plugin marketplace add microsoft/azure-skills, then browse /plugins and install azure. See Codex skills. For Cursor, the README points to the Cursor Marketplace or Settings, then Plugins; see Cursor skills.
Skill content only. Learn says to use this if your host already has MCP server configuration:
npx skills add https://github.com/microsoft/azure-skills --skill microsoft-foundry
Prerequisites, per Learn: an Azure subscription, Node.js 18 or later (the plugin uses npx to start MCP servers), Git, the Azure CLI signed in with az login, the Azure Developer CLI signed in with azd auth login for azd workflows, and a Foundry project for project-scoped work.
How do you use it?
Open the folder with your agent code and ask for the outcome. Learn says you don't usually need to name the sub-skill; include the project, environment, agent folder or deployment name if you know them.
Example prompts (suggestions, adapted from the patterns on Microsoft Learn):
Use the Microsoft Foundry Skill to create a public Foundry project, deploy a model, and scaffold a hosted agent.Prepare this hosted agent for deployment to my dev environment and smoke-test it.Create an evaluation plan for this agent from eval.yaml, then run a batch eval.Check RBAC and quota blockers before I deploy this model.Troubleshoot why this hosted agent deployment is failing, then redeploy and retest.
To check the setup, Learn suggests asking What AI models are available in Microsoft Foundry? and List my Azure resource groups. Both should come back from real tools, not a generic summary.
Tips and limits
From Microsoft's docs, the repo and skills.sh, not from testing:
- It acts on real, billable resources. Learn says: "For deployment, fine-tuning, and provisioning tasks, confirm role assignments and cost-bearing resources before the agent runs commands." Hosted agent runtime is billed on CPU and memory used in active sessions, per the hosted agents page.
- It installs an azd extension. The first-step script runs
azd extension install microsoft.foundrywhen the extension is missing. - The plugin adds telemetry hooks. The Azure plugin's Claude Code hooks run a telemetry script on session start and after tool use. The README says to set
AZURE_MCP_COLLECT_TELEMETRY=falseto opt out, and the hook script checks that variable. - Private networks block MCP. The SKILL.md says "Foundry MCP tools cannot reach a VNet-isolated project even from inside the VNet."
- Don't store secrets in
.foundryfiles. Learn says to keep only nonsecret overlay state there. - Security audit. On skills.sh, Socket rated it Warn (audited 15 October 2026). It flagged a fine-tuning script that runs a grader file with
exec()in the current Python process (medium) and an example that trusts a public unauthenticated MCP server with approval turned off (low), and reported no evident malware. Gen Agent Trust Hub and Snyk rated it Pass.
The MIT License lets you copy and change the skill as long as you keep Microsoft's copyright notice.
Save it in your Kipt library
You can keep a copy of microsoft-foundry in Kipt: open Skills and paste the skill's GitHub folder link into the import box, or drop the folder or a .zip. A copy saves the skill files only. The Azure MCP Server configuration and telemetry hooks are part of the plugin, not the skill folder, so use the plugin route if you need the MCP tools wired up. Skills 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.
Similar skills
- Azure diagnostics skill: Microsoft's skill for troubleshooting Azure production issues, from the same plugin.
- Claude API skill: Anthropic's skill for building on the Claude API.
- Hugging Face skill: the Hugging Face CLI skill for models and datasets on the Hub.
- MCP builder skill: Anthropic's guide to building your own MCP server.
- Best MCP servers: a list of MCP servers to pair with skills.