MCP vs Skills
When to Use Which
Matthew Khouzam | Ericsson Research | May 2026
This talk covers the two major paradigms for extending AI agent capabilities: MCP (Model Context Protocol) for tools and data access, and Skills for instructions and workflows. We'll explore when to use each, and how they complement each other.
What They Are
Let's start with definitions of both technologies.
MCP (Model Context Protocol)
Open standard (Nov 2024), donated to Linux Foundation's Agentic AI Foundation (Dec 2025)
Universal way to give LLMs tools and real-time data access
Client-server architecture: AI app → MCP client → MCP server → tools/resources
MCP provides a universal protocol layer so that any AI client can connect to any tool server.
MCP: Adoption
97M+ monthly SDK downloads, 10,000+ active servers
Backed by Anthropic, OpenAI, Google, Microsoft, AWS
"The USB-C port of AI applications" — Anthropic
Think of it like LSP for IDEs but for AI agents. Massive industry adoption.
Skills
Markdown-based instruction sets stored as directories on the filesystem
Core: a SKILL.md file with YAML frontmatter, optional scripts, docs, references
Introduced by Anthropic (Oct 2025)
Skills are plain English instructions that teach the AI how to do something. No code required.
Skills: How They Load
Progressive disclosure : metadata at startup (~100 tokens) → full body when relevant (~5,000 tokens) → additional files on demandWorks across Claude.ai, Claude Code, Agent SDK, Claude API
A skill tells the agent how to approach a task — like a style guide handed to a developer
Progressive disclosure keeps context costs low — only what's needed gets loaded.
Key Differences
Let's compare MCP and Skills across multiple dimensions.
Comparison: What & How
Dimension MCP Skills Nature Tools and data access Instructions and workflows How they work Called on-demand via tool calls Loaded into context progressively Data Real-time, live (APIs, DBs, tickets) Static reference material
The fundamental difference: MCP gives the AI new abilities to reach external systems. Skills give the AI knowledge about how to do things.
Comparison: Scope
Dimension MCP Skills Scope External systems, network resources Local filesystem, coding patterns Auth Often requires tokens/OAuth None needed
MCP reaches out to external systems and often needs credentials. Skills stay local.
Comparison: Creation
Dimension MCP Skills Creation Requires writing code (server impl) Plain English markdown files
MCP requires engineering effort. Skills are just markdown files anyone can write.
Comparison: Context Cost
Dimension MCP Skills Context cost Progressive discovery (implemented in some agents, not yet officially in the spec) Progressive disclosure (implemented in some agents, not yet officially in the spec)
Progressive discovery for MCP and progressive disclosure for Skills are implemented in some agents (e.g. Claude Code, Kiro) but neither is officially part of the respective specs yet.
Comparison: Portability & Distribution
Dimension MCP Skills Portability Any AI app with MCP client Primarily Claude ecosystem Distribution Hosted servers (local or cloud) Files in a directory (git)
MCP is more portable across AI vendors. Skills are easier to distribute via version control.
Comparison: Lifecycle
Dimension MCP Skills Persistence Only active when connected Always active once installed Analogy Power tools on a workbench Style guide for a developer
MCP tools disappear when the server disconnects. Skills persist as long as they're in the project.
When to Use MCP
MCP is the right choice when you need to reach out to external systems.
MCP Use Cases (1/2)
Live data access — current inventory, real-time API responses, database queries, latest emailsExternal system interaction — creating GitHub PRs, reading Jira tickets, querying Figma, deploying resourcesCross-model compatibility — build tool integrations once for multiple AI models
MCP is for when the AI needs to reach outside its sandbox to interact with the real world.
MCP Use Cases (2/2)
Authenticated operations — anything requiring API keys, OAuth tokens, credentialsSimple data retrieval — one API call gets what you need
Authentication and simple retrieval are classic MCP territory — the AI can't do these without external access.
When to Use Skills
Skills are the right choice when you need to teach the AI how to do something.
Skills Use Cases (1/2)
Complex multi-step workflows — financial modeling, compliance reviews, data analysis pipelinesOrganizational knowledge — brand guidelines, SOPs, domain methodologies, coding conventionsTeam standardization — commit skills to repo; every dev gets identical AI behavior
Skills encode knowledge and procedures. They're about the how, not the what.
Skills Use Cases (2/2)
Framework best practices — "prefer Server Components," "always use named exports"Deterministic operations — data validation, calculations, file format conversionsMulti-tool orchestration — defining the order and logic of how tools combine
Skills can also encode executable logic and orchestration patterns, not just prose instructions.
The Sweet Spot: Using Both
MCP and Skills are complementary, not competing.
The Mental Model
MCP gets data to the AI. Skills tell the AI what to do with it.
Skills encode how to do something.ability to do something the AI couldn't do before.
This is the key insight. They solve different problems and work best together.
Examples: Both Together
Notion : MCP server provides 15+ CRUD tools → Skill defines how to create a specific page type (sections, format, research steps)Deployment : Skill defines the process ("check CI → merge → smoke tests → promote") → GitHub & Playwright MCP servers execute each stepSales CRM : MCP connects to transcription service & CRM API → Skill encodes logic for extracting key points and mapping to CRM fields
In each case, MCP provides the hands and Skills provide the brain. The Skill knows what to do; MCP knows how to reach the external systems.
Anti-Patterns to Avoid
Common mistakes when choosing between MCP and Skills.
Don't Do This (1/2)
❌ Putting API docs in a skill — they go stale; use a live docs MCP server
❌ Using MCP to enforce coding conventions — use a skill + linter
❌ Installing too many skills — context window bloat; prioritize 5–10
The key principle: don't use one where the other is a better fit. Static knowledge goes in Skills, live access goes in MCP.
Don't Do This (2/2)
❌ Duplicating MCP tool instructions in a skill — the server already knows how to create a PR; the skill should describe your team's PR conventions
❌ Too many MCP tools the LLM doesn't need — causes "context rot" degrading performance
Redundancy between MCP and Skills wastes context and confuses the model.
What About Direct API Access?
Some teams skip MCP and hardcode API calls. Let's discuss why that's usually a mistake for agentic workflows.
Direct API: The Pre-LSP World
Direct API
Bound to a single tool
Only works in one agent
No discoverability
No standard schema
No reuse across models
Bespoke, brittle, duplicated
MCP
Universal protocol layer
Works across any AI client
Self-describing tools
Model reasons dynamically
Build once, use everywhere
Like LSP for AI agents
Before LSP, every editor wrote its own integration for every language. LSP defined clear roles and created a vibrant ecosystem. MCP does the same for AI agents.
Direct API access still makes sense for one-off scripts or tightly controlled pipelines, but for agentic workflows it quickly becomes a maintenance burden.
Where Things Are Heading
Both technologies are evolving rapidly.
Future Directions (1/2)
Skills gaining traction faster than MCP did at the same stage
MCP evolving: streamable HTTP , OAuth 2.1 , headless microservice auth
MCP now vendor-neutral under Linux Foundation governance
Both technologies are maturing rapidly with strong industry backing.
Future Directions (2/2)
MCP Apps (Jan 2026): tools return interactive UI components in conversationFor local coding agents, skills can substitute for some MCP via CLI-calling scripts
For headless cloud microservices , MCP remains essential — skills don't apply
The two technologies are converging in some areas but remain distinct in their core value propositions.
Enterprise Evidence
Rakuten : ~87.5% faster financial reporting with SkillsE-commerce company: automated 50%+ of support tickets, saving ~$2M/year
Python API for programmatic skill use in custom agents (early-stage)
Real enterprise results showing the value of both approaches.
Rule of Thumb
If it needs to know how → Skill
If it needs to reach out → MCP
If it needs both → use both
This is the simplest decision framework. Know how = Skill. Reach out = MCP. Both = both.
Quiz: MCP or Skill?
For each scenario, decide: MCP, Skill, or Both?
Interactive quiz to test understanding.
Quiz (1/3)
# Scenario Answer 1 Always write React components using Server Components + named exports Skill 2 Read Jira tickets and create GitHub PRs that reference them MCP 3 12-step deployment checklist: check CI, smoke tests, promote to prod Both 4 Format every Excel report with company header, colors, pivot layout Skill 5 Query Postgres for latest sales figures MCP
1: coding conventions = Skill. 2: live API access to two systems = MCP. 3: Skill defines procedure, MCP servers execute. 4: deterministic formatting = Skill. 5: real-time data = MCP.
Quiz (2/3)
# Scenario Answer 6 All devs on team get same AI coding behavior when they clone the repo Skill 7 Post a message to Slack when a build fails MCP 8 Complex financial model: variance, anomalies, formatted PDF Skill 9 Fetch latest docs for a frequently-updated library MCP 10 Structure PR descriptions with summary + test plan, then create the PR Both
6: commit to repo = Skill. 7: authenticated Slack API = MCP. 8: multi-step procedural workflow = Skill. 9: live docs = MCP. 10: Skill for template, MCP to create.
Quiz (3/3)
# Scenario Answer 11 Authenticate with OAuth and access Google Drive files MCP 12 Analyze call transcripts and map info into CRM fields per team taxonomy Both 13 Enforce parameterized SQL queries, never SELECT * Skill 14 Check real-time inventory across three warehouses MCP 15 Generate brand-compliant marketing emails in company voice Skill
11: credential handling = MCP. 12: MCP for CRM + transcription, Skill for mapping logic. 13: coding standard = Skill. 14: live data = MCP. 15: brand guidelines = Skill.
References
Shaw, "Agent Skills vs MCP: What's the difference?" — YouTube
Sterling, "Claude Skills vs MCP: What's the Difference and When to Use Each?" — YouTube
"Agent Skills or MCP in the era of Claude Code?" — YouTube
Developer Toolkit, "Skills vs MCP — When to Use Which" — developertoolkit.ai
Anthropic, "Skills explained" — claude.com/blog
IntuitionLabs, "Claude Skills vs. MCP: A Technical Comparison" — intuitionlabs.ai
Full URLs available in the companion markdown document.
Thank You
Matthew Khouzam | Ericsson Research
May 2026
