Microsoft Foundry

Foundry-platform fluency for any language. Skills teach the model the mental model + non-obvious gotchas, then direct it to discover current API surface via Foundry MCP (mcp.ai.azure.com), Microsoft Docs MCP, and azd ai agent. No baked-in stale SDK syntax — concepts in, fresh code out.

Install

# Copilot CLI
/plugin marketplace add microsoft/skills
/plugin install microsoft-foundry@skills

# npx
npx skills add microsoft/skills --skill microsoft-foundry

What You Get

MCP Servers

The plugin bundles three MCP servers that activate automatically on install:

Skills

One orchestrator routes intent to ten language-agnostic sub-skills. Each sub-skill teaches the platform mental model and gotchas, then directs to the appropriate discovery surface for current syntax.

Companion Skills

Foundry agents commonly call into adjacent Azure SDKs. Those skills live in their language plugins — install the matching azure-sdk-* plugin for SDK-specific guidance:

This plugin stays Foundry-platform focused; install language plugins alongside it for SDK depth.

Prerequisites

• Azure account with an AI Foundry project (create one)
• Azure CLI authenticated (az login)
• Node.js 18+ (for the Azure MCP server)
• azd extension for hosted agents: azd ext install azure.ai.agents

Authentication

Foundry uses Microsoft Entra. Authenticate via Azure CLI:

az login                  # for az + Azure MCP + DefaultAzureCredential locally
azd auth login            # for hosted-agent deploys via `azd ai agent`

In code, prefer DefaultAzureCredential for local dev and ManagedIdentityCredential in production (hosted agents, App Service, Container Apps, AKS).

Set your Foundry project endpoint:

export AZURE_AI_PROJECT_ENDPOINT="https://<resource>.services.ai.azure.com/api/projects/<project>"

Quick Start

After installing the plugin, your coding agent routes intent through the orchestrator to the right sub-skill:

"Provision a Foundry project and a model deployment"
  -> foundry-projects-resources + foundry-models

"Deploy a hosted agent with the Responses protocol"
  -> foundry-hosted-agents (azd ai agent + agent.yaml)

"Curate a toolbox with Web Search + AI Search + an MCP server"
  -> foundry-toolboxes (single MCP endpoint, versioned)

"Ground my agent on our SharePoint knowledge with permission trimming"
  -> foundry-iq-knowledge-bases (agentic retrieval + ACL-aware)

"Set up tracing, batch evals, and a regression alert"
  -> foundry-observability (App Insights KQL + evaluators)

"Centrally govern which MCP tools my agent fleet can call"
  -> foundry-governance (Foundry Control Plane + AI Gateway)

Plugin Structure

microsoft-foundry/
  .claude-plugin/
    plugin.json                       # Plugin manifest
  .mcp.json                           # Azure MCP + Foundry MCP + Microsoft Docs MCP
  README.md
  skills/
    microsoft-foundry/                # Orchestrator — routes intent to sub-skills
    foundry-projects-resources/       # Resources, projects, connections, auth
    foundry-models/                   # Deploy, capacity, quota, RAI policy
    foundry-hosted-agents/            # Responses + Invocations, agent.yaml, azd
    foundry-toolboxes/                # Intent-based toolboxes (9 tool types)
    foundry-workflows/                # Workflow vs A2A vs Connected Agents
    foundry-iq-knowledge-bases/       # Foundry IQ + agentic retrieval
    foundry-managed-skills/           # SKILL.md as a Foundry resource
    foundry-memory/                   # Long-term memory stores
    foundry-observability/            # Tracing + KQL + evals + RAI evals
    foundry-governance/               # Control Plane + AI Gateway + RBAC

Skills are symlinked to canonical sources in the azure-skills plugin to maintain a single source of truth.

Doctrine: Concepts → Discover

Foundry surfaces evolve fast. Skills that bake in SDK syntax go stale and produce broken code.

This plugin takes the opposite approach:

• Concepts in skills. Each sub-skill teaches the mental model (what a toolbox is, why memory has prompt-injection risk, how Responses differs from Invocations) and the non-obvious gotchas (preview header flags, capacity-region mismatches, ACL trimming caveats).
• Discover for syntax. Skills direct the model to the live surface for current API shape:
• Microsoft Docs MCP — official Learn docs, always fresh
• Foundry MCP (mcp.ai.azure.com) — Foundry-native operations
• azd ai agent — hosted agent build/deploy/monitor
• az CLI — Azure resource management

The result: stable skills + always-current code.

License

MIT