MicrosoftDocs/mcp — Summary

Microsoft Learn MCP Server is Microsoft's official documentation-retrieval MCP server — a free, no-auth, remote HTTP MCP endpoint (https://learn.microsoft.com/api/mcp) that gives AI agents real-time access to the complete Microsoft documentation corpus. It ships 3 MCP tools, 3 Agent Skills, a TypeScript CLI (mslearn), and multi-IDE support (Claude Code plugin, Codex plugin, .agents configuration).

Problem it solves: AI models hallucinate Microsoft API methods, invent Azure SDK parameters, and use outdated patterns because their training data is stale. The Learn MCP server provides live, authoritative, first-party documentation retrieval that eliminates these hallucinations.

Distinctive trait: The only documentation-specialized MCP server in the corpus from a primary vendor. It ships BOTH an MCP server AND Agent Skills that guide agents on when to use each tool — a unique "MCP server + skills" combination where the skills act as smart routing instructions above the MCP layer. The microsoft-skill-creator skill adds meta-skill creation capability.

Target audience: Any AI developer using Microsoft technology — Azure, .NET, M365, Power Platform, Windows — who wants their coding agent to have accurate, up-to-date Microsoft documentation access.

Scope: 1,661 GitHub stars, CC-BY-4.0 license (for content), TypeScript, 195 forks, active maintenance.

Differs from seeds: Unlike taskmaster-ai (Archetype 3 — MCP-anchored for task management), ms-learn-mcp is documentation-specialized MCP — the MCP server is a read-only knowledge retrieval system, not a task orchestrator. The skills layer on top of the MCP server is what separates this from a raw API — skills teach agents WHEN to use microsoft_docs_search vs microsoft_code_sample_search, implementing skill-as-router pattern.

MicrosoftDocs/mcp — Overview

Origin

Published by MicrosoftDocs GitHub organization — the same org that maintains Microsoft's official documentation. This gives the server uniquely authoritative access to its own documentation corpus.

Philosophy

"Stop AI Hallucinations. Give your AI assistant (Claude, Cursor, Copilot, Codex, ...) direct access to the latest official Microsoft documentation." "✨ Free. One-click install. No key needed."

Key principles:

1. Trusted source over web scraping: "Unlike generic web searches that may scrape insecure blogs or malicious sites, this tool only accesses official 1st-party Microsoft documentation"
2. Zero friction: No API keys, no registration, no credits — just install and use
3. Recency: Live access to current docs beats training data cutoff
4. Hallucination prevention: Code that compiles vs. code that invents non-existent SDK methods

The Hallucination Problem

The README calls out three specific hallucination scenarios:

• "AI inventing non-existent Azure SDK methods"
• "Hallucinating library packages"
• Wrong method signatures causing compilation failures

Multi-Client Design

The README explicitly supports all major AI clients:

• Claude (via Claude Code plugin and .claude-plugin/)
• Cursor
• GitHub Copilot
• Codex (via .codex-plugin/ — note the .codex-plugin in repo)

Experimental Features

Two experimental features indicate roadmap direction:

1. OpenAI Deep Research compatibility (/api/mcp/openai-compatible) — targeting OpenAI's research model format
2. Token budget control (?maxTokenBudget=2000) — managing cost for heavy usage sessions

MicrosoftDocs/mcp — Architecture

Distribution

• Type: mcp-server + skill-pack + CLI tool
• Primary endpoint: https://learn.microsoft.com/api/mcp (remote, free, no auth)
• License: CC-BY-4.0 (content), custom for code
• Protocol: Streamable HTTP

Repository Structure

MicrosoftDocs/mcp/
├── .agents/                        # Agent configuration files
├── .claude-plugin/                 # Claude Code plugin manifest
├── .codex-plugin/                  # Codex plugin manifest (note: uses .codex-plugin)
├── .mcp.json                       # MCP server config
├── .openpublishing.publish.config.json  # Publishing config
├── cli/                            # TypeScript CLI (@microsoft/learn-cli)
│   ├── src/                        # CLI source
│   ├── test/
│   ├── package.json
│   └── README.md
├── scripts/                        # Build/publish scripts
├── skills/                         # 3 Agent Skills
│   ├── microsoft-docs/
│   │   └── SKILL.md               # Documentation skill (concept understanding)
│   ├── microsoft-code-reference/
│   │   └── SKILL.md               # Code sample + API lookup skill
│   └── microsoft-skill-creator/
│       └── SKILL.md               # Meta-skill for creating more skills
├── AGENTS.md
├── README.md
├── SECURITY.md
└── ThirdPartyNotices.md

MCP Server Configuration (.mcp.json)

{
  "servers": {
    "microsoft-learn": {
      "type": "http",
      "url": "https://learn.microsoft.com/api/mcp"
    }
  }
}

MCP Tools (3)

CLI Package

• NPM: @microsoft/learn-cli
• Binary: mslearn
• Instant run: npx @microsoft/learn-cli search "azure functions timeout"

Required Runtime

• No runtime for the remote MCP endpoint
• node >= 22 for the TypeScript CLI

MicrosoftDocs/mcp — Components

MCP Tools (3)

Agent Skills (3)

microsoft-docs (SKILL.md)

Purpose: Understanding concepts, tutorials, architecture, limits

When to activate: User asks "how does X work", wants tutorials, needs config options, limits/quotas, best practices for any Microsoft technology. Activates even without mentioning "docs."

Key routing logic:

• microsoft_docs_search → find documentation
• microsoft_docs_fetch → full page content when excerpts insufficient
• Falls back to mslearn CLI if MCP server unavailable

microsoft-code-reference (SKILL.md)

Purpose: Find working code samples, verify API signatures, fix Microsoft SDK errors

When to activate: User is writing, debugging, or reviewing code touching any Microsoft SDK — catches hallucinated methods, wrong signatures, deprecated patterns.

Key routing logic:

• microsoft_code_sample_search → working code patterns
• microsoft_docs_search → API method/class lookup
• microsoft_docs_fetch → full API reference pages

microsoft-skill-creator (SKILL.md)

Purpose: Meta-skill for creating additional Microsoft Learn skills

Key feature: Enables users to create custom skills that teach agents to use the Learn MCP tools for specific Microsoft products/scenarios — expanding the skill library.

CLI Commands (mslearn)

IDE/Agent Manifests

MicrosoftDocs/mcp — Uniqueness & Positioning

Differs from Seeds

Unlike all 11 seeds (workflow/skill enforcement frameworks), ms-learn-mcp is a documentation retrieval service with a skills layer on top. The closest analogy is ccmemory (Archetype 3 — MCP-anchored toolserver for memory) but MicrosoftDocs/mcp targets documentation retrieval, not memory storage.

The key architectural innovation: MCP server + Agent Skills combination — the 3 skills act as routing instructions above the 3 MCP tools, implementing a "smart proxy" pattern. The agent doesn't pick tools directly; skills determine the right tool based on the user's intent. This separation of "what to use" (MCP tools) from "when and how to use it" (skills) is not present in any seed.

Distinctive Properties

1. First-party documentation authority: The MicrosoftDocs GitHub organization builds both the documentation AND the MCP server that retrieves it — unmatched domain authority
2. Zero-friction adoption: Free, no auth, one-click install — removes every barrier to adoption
3. Skills-as-router pattern: 3 skills route to 3 tools based on intent — a routing layer that no pure MCP server implements
4. context: fork extension: A documented extension to the agentskills.io spec — Microsoft extends the standard for context isolation needs
5. Multi-IDE with Codex support: The .codex-plugin/ directory signals early support for OpenAI's Codex agent format — broadest IDE coverage in the batch
6. OpenAI Deep Research compatibility: Experimental /openai-compatible endpoint for OpenAI research models — cross-vendor interoperability beyond just MCP
7. Token budget control: Server-side truncation parameter for context budget management — addresses agent context window limits directly in the API

Observable Failure Modes

• CC-BY-4.0 license restriction: Content license may create friction for commercial use of documentation excerpts
• Single source dependency: All retrieval goes through learn.microsoft.com/api/mcp — if the endpoint has downtime, all retrieval fails (mitigated by CLI fallback)
• Microsoft-only scope: Zero value for non-Microsoft technologies — the narrowest knowledge scope of any "documentation" MCP server
• Skills lag spec changes: As agentskills.io spec evolves, the context: fork extension may conflict with standard behavior

Cross-References in Batch

• Implements agentskills.io format (see asyncope-agentskills-io) with context: fork extension
• Complementary to ms-mcp-servers (Azure MCP covers operations; Learn MCP covers documentation)
• Referenced by google/agents-cli README as an example of skills-compatible agent tooling