MCP & major servers

1. What MCP is

Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems. Think USB-C: one unified interface, so you don't write a bespoke integration for every connection target (Slack / GitHub / Drive / internal DB / …).

Transports

• stdio: local execution (Claude Code, Desktop, VS Code, Cursor). A server process runs on the user's machine and communicates via stdin/stdout.
• HTTP: remote server (Claude API MCP Connector). Shared cloud-hosted server, connected over HTTP/SSE — usable from multiple clients.

Capability negotiation

On connect, the server declares the resources (files, DBs), tools (APIs, scripts), and prompts it exposes. The client (Claude) only invokes what it actively needs — which keeps context size bounded instead of ballooning.

Security boundaries

• Connection control: the client manages which servers are allowed to connect (allow-list).
• Exposure scope: the server chooses which resources and tools to expose.
• Encryption: all traffic is TLS. HTTP transport supports auth headers and OAuth.

2. MCP server catalog

The official MCP Registry (registry.modelcontextprotocol.io) lists 110+ servers — a mix of Anthropic-official and community implementations.

Anthropic official servers

Server	Purpose
Filesystem	Read / write local files
Git	Git repository operations
GitHub	GitHub Issues / PRs / Actions
Google Drive	Search and fetch files in Drive
Slack	Channels, messages
Notion	Notion databases / pages
Linear	Issue tracking
Figma	Design file reference
Postgres	SQL execution
BigQuery	Data queries
Memory	Long-term memory (knowledge graph)
Fetch	URL fetch
Sequential Thinking	Chain-of-thought
Time	Clocks / time zones

Community servers

90+ additional implementations come from the community. Searchable via third-party indexes like Glama Registry (glama.ai). Strong coverage of business SaaS — Salesforce, Jira, Confluence, Zendesk, HubSpot, and more.

3. Five executive workflows

Five practical, non-engineer workflows that become possible when executives themselves use Claude + MCP.

Workflow 1 · Slack triage

MCP Slack server → auto-generated morning Daily Digest per channel. Unread messages classified by priority; only the reply-worthy ones piped to email or a dashboard.

Workflow 2 · Drive audit

MCP Google Drive → auto-log new and changed files from the last 7 days into Notion. Makes "who changed what when" visible; semi-automates access audits on confidential material.

Workflow 3 · GitHub roll-ups

MCP GitHub → weekly progress report (merged PRs, open issues, CI failure rate). Collapses the Monday-morning prep time for 1:1s and executive reports nearly to zero for engineering managers.

Workflow 4 · Calendar integration

Multiple MCP servers in parallel (Google Calendar + Gmail + Slack) → auto-delivered "Executive Summary: your week" every Monday morning. Key meetings, prep materials, critical unread messages — consolidated to one screen.

Workflow 5 · Linear dashboard

MCP Linear → real-time project visibility. Blocker detection, effort-estimate auto-updates, early warnings on slip risk — surfaced as a dashboard.

4. The Skills system

Agent Skills are reusable modules. If MCP handles "external system connectivity," Skills handle "repeated workflow guidance" — complementary layers. Every Skill is a single file: SKILL.md (YAML frontmatter + Markdown instructions).

SKILL.md structure

---
name: skill-name
description: what it does + when to use it
license: MIT (optional)
allowed-tools: Bash(git:*) Read (optional)
---

# Detailed instructions

Progressive disclosure

Skills use three-stage progressive loading so context isn't wasted:

1. Metadata (~100 tokens): always loaded at startup. Claude only needs to know the name and description.
2. Instructions (<5,000 tokens): loaded when the skill activates. Detailed runtime steps.
3. Resources (variable): scripts/, templates, references — pulled only when needed.

Effect: you can install 100+ skills and still keep context usage minimal.

agentskills.io standard

The industry-common spec announced by Anthropic in December 2025. Adopted by 26 tools (Cursor, GitHub Copilot, Claude Code, Gemini, Codex, Windsurf, …). One SKILL.md becomes portable across multiple AI tools.

5. Official Anthropic Skills

Main Skills published by Anthropic (github.com/anthropics/skills):

• frontend-design: frontend design generation
• systematic-debugging: systematic debugging
• brainstorming: brainstorming
• writing-plans: implementation planning
• executing-plans: plan execution
• test-driven-development: TDD
• subagent-driven-development: subagent-driven development
• requesting-code-review: requesting a code review
• verification-before-completion: verify-before-done
• finishing-a-development-branch: closing out a dev branch
• receiving-code-review: receiving a review
• using-git-worktrees: using git worktrees
• using-superpowers: Superpowers skills meta-guide
• mcp-builder: building MCP servers
• skill-creator: skill authoring helper

Plus Pre-built Document Skills (PowerPoint, Excel, Word, PDF) usable on Desktop / Claude.ai.

6. Creating a skill

A skill is defined by a single Markdown file. The create-to-share flow:

1. mkdir ~/.claude/skills/my-skill — create the directory
2. Write SKILL.md (frontmatter + instructions)
3. Hot reload: file changes auto-detected, no Claude restart needed
4. Packaging: bundle as <plugin>/skills/<name>/SKILL.md
5. Custom marketplace: point settings.json at a GitHub repo URL → share skills internally

7. Skills vs subagents vs slash commands

Claude Code exposes three customization mechanisms with clearly different shapes.

Trait	Skill	Subagent	Slash command
Auto-activation	On description match	Manual only	Manual only
Context	Continues current thread	Isolated (new context)	Continues current thread
Use case	Workflow guidance	Deep-dive research	One-shot operation
Complexity	Low to medium	High	Low
Distribution	plugin / marketplace	Project-local definition	CLAUDE.md

Picking between them

• Repetitive workflow → Skill (e.g., weekly report creation)
• Independent deep-dive task → Subagent (e.g., long research runs, high-volume file analysis)
• One-shot simple command → Slash command (e.g., /deploy, /status)

8. Availability by surface

Which Claude surface you use determines what Skills and MCP you can run.

Capability	Claude Code	Desktop	Claude.ai	Mobile
Pre-built Skills	✗	✓	✓	✗
Custom Skills	✓	✓	✓ (Pro+)	✗
MCP (stdio)	✓	✓	✗	✗
MCP Connector (HTTP)	✓	✗	✗	✗

9. Sharing & discovery

Skills and MCP servers distribute through two layers: the official marketplace and a company-internal custom marketplace.

Official marketplace

Anthropic's official skills install directly from inside Claude Code:

/plugin install document-skills@anthropic-agent-skills

Custom marketplace (internal)

Register your company's GitHub repository as a marketplace — every employee then runs the same command to pull internal skills:

{
  "extraKnownMarketplaces": {
    "my-org": {
      "source": { "source": "github", "repo": "myorg/skills" }
    }
  }
}

→ /plugin install my-feature@my-org pulls internal skills.

10. Official sources

Primary references: MCP spec, server registry, Skills spec, Anthropic's official Skills collection, the Claude API / Claude Code Skills guides, and Anthropic Engineering blog posts. Pull from these at each validation stage of rollout. Aggregated in the Sources block below.

This page reflects information as of 2026-04-23. The MCP registry has grown multiple-fold in the year since launch — re-check registry.modelcontextprotocol.io when the date matters.