Microsoft Learn MCP Server

The Microsoft Learn MCP Server is Microsoft's official Model Context Protocol server for its own technical documentation. According to the official repository README, it addresses a very concrete problem: AI coding assistants occasionally invent non-existent Azure SDK methods or hallucinate package names because their training knowledge is outdated. Instead of relying on that frozen knowledge or on risky, uncontrolled web searches, the server connects compatible clients directly to current, official Microsoft documentation. The official repository lives at https://github.com/MicrosoftDocs/mcp (exactly 1,898 stars per the GitHub API on 2026-09-22, actively maintained, last pushed 2026-09-10), with the corresponding product documentation at https://learn.microsoft.com/training/support/mcp.

What is the Microsoft Learn MCP Server?

Technically, this is a hosted remote MCP server operated by Microsoft itself at https://learn.microsoft.com/api/mcp, reachable via Streamable HTTP. According to the provider, it runs on the same RAG-based "Ask Learn" knowledge service that also powers Copilot for Azure. There is no local installation path in the classic sense: the server runs exclusively remotely, and no API key or Microsoft account is required according to the official product page ("There's no authentication required to access the Microsoft Learn MCP Server"). This sets it apart clearly from the other two Microsoft entries in the Skill Road catalog: the Azure MCP Server manages concrete, account-bound Azure resources, and the Fabric MCP Server operates against Microsoft Fabric workspaces — both require Azure sign-in. The Learn MCP Server, in contrast, only reads publicly accessible documentation content and never modifies any account or resource.

Features and tools

According to the README, the server exposes three tools: microsoft_docs_search for semantic search across the entire official technical documentation, microsoft_docs_fetch for retrieving and converting a specific documentation page into Markdown, and microsoft_code_sample_search for targeted search of official Microsoft and Azure code samples, optionally filtered by programming language. This allows questions such as "Give me the Azure CLI commands to create an Azure Container App with a managed identity" or "Is GPT-5.4 available in Azure EU regions?" to be answered using current, official content instead of frozen model knowledge. For experimental use cases, the README describes an additional OpenAI-compatible endpoint at /api/mcp/openai-compatible for OpenAI Deep Research models, as well as a maxTokenBudget query parameter that limits response length for cost control. Both extras are described by the provider as still under active development and subject to change.

CLI tool and Agent Skills

Alongside the MCP endpoint, Microsoft offers @microsoft/learn-cli (npm, MIT licensed, current version 1.0.0 per the npm registry), a command-line tool that exposes the same search functionality without any MCP client, for example via npx @microsoft/learn-cli search "azure functions timeout". The repository also ships three portable Agent Skills (microsoft-docs for concepts and tutorials, microsoft-code-reference for API lookups and troubleshooting, microsoft-skill-creator as a meta-skill that generates custom skills for specific Microsoft technologies), giving agents extra guidance on when and how to use the MCP tools effectively.

Who should use it?

The server is especially useful for teams working regularly with .NET, Azure, Microsoft 365, or other Microsoft technologies who need reliable, current information in their AI development environment instead of outdated training data. Because no authentication or account is needed, the barrier to entry is minimal — a single URL entry is enough in most clients. Anyone who actually needs to manage Azure resources or edit Fabric workspaces still needs the account-bound servers already listed in the catalog.

Setup

According to the README, the server supports virtually every common MCP client with one-click installation or a few lines of configuration: VS Code and Cursor offer direct install buttons, Codex just needs codex mcp add "microsoft-learn" --url "https://learn.microsoft.com/api/mcp", and Claude Code has an official plugin (/plugin install microsoft-docs@claude-plugins-official) that installs both the MCP server and the Agent Skills together. Clients without native remote MCP support can add the standard configuration manually using "type": "http" and the URL above.

Security notes

Because the server, according to the provider, only offers read access to publicly published Microsoft documentation, requires no authentication, and processes no account or customer data, its risk profile is very low. There are no write operations and no credentials that could be compromised. The manufacturer itself notes in the README that MCP is an evolving protocol: anyone building a custom client should discover tools dynamically via tools/list rather than hard-coding them, since tool definitions can change.

FAQ

Is the Microsoft Learn MCP Server free? Yes. According to the official product page, the server is publicly available and there is no charge to use it.

Do I need a Microsoft or Azure account? No. Unlike the Azure MCP Server or Fabric MCP Server in the catalog, the documentation search needs no account and no authentication.

How does it differ from the Azure MCP Server? The Azure MCP Server manages account-bound Azure resources and requires sign-in. This Learn server only provides read access to public documentation and code-sample content, with no sign-in at all.

A compatible MCP client with remote MCP server support (Streamable HTTP); no Microsoft or Azure account or authentication required.

For Claude Code: /plugin install microsoft-docs@claude-plugins-official (installs both the MCP server and the Agent Skills together, then restart Claude Code). For Codex: codex mcp add "microsoft-learn" --url "https://learn.microsoft.com/api/mcp". In VS Code and Cursor, the one-click install button from the official README is enough. Other clients can add the standard configuration manually using "type": "http" and the server URL.

codex mcp add "microsoft-learn" --url "https://learn.microsoft.com/api/mcp"

No authentication required — the server is publicly reachable and needs no Microsoft or Azure account.

Read-only access to public Microsoft developer documentation and official code samples; no access to account, subscription, or resource data.

Requests go through the Microsoft-hosted Streamable HTTP endpoint at learn.microsoft.com/api/mcp; search internally uses the same RAG-based knowledge service that powers Ask Learn and Copilot for Azure. Responses are returned to the MCP client and its language model.

Very low risk profile: read-only access to public content, no authentication, no account or customer data involved. Retrieved documentation excerpts flow into model requests.