Skills vs MCP Servers: when to pick which

Giving your Claude Code workspace superpowers

Around $285 billion in legacy enterprise software market cap evaporated last week.

The catalyst wasn’t a recession or a scandal.

It was a small free software repository that Anthropic made available.

In the repository were eleven industry-specific plugins for Claude that covered workflows in legal, financial modeling, marketing, sales and customer service.

This small repository caused quite the ruckus:

Salesforce lost 14% in 5 days. SAP around 16%. Thomson Reuters shares fell 18%.

In total, the S&P 500 software index dropped nearly 9% over five trading sessions.

Investors dubbed it the “SaaSpocalypse.”

But as regular readers will know, the plugins in that repository aren’t anything special.

They’re simple bundles of Skills and MCP server connections—which are available to anyone running Claude Code.

Given the market response, I figured it’d be good to take another look at what Skills and MCP servers are, and how to best use them.

Because there are some key differences between Skills and MCP servers, and knowing which one to use when will save you a lot of wasted effort.

Skills: teaching your AI how to work

A Skill is a set of instructions for AI models written in plain text.

These instructions are stored in a SKILLS.md file in a specific location so your AI agent knows where to look for them.

Here’s an example Skills definition:

---
name: summarize-research
description: Summarize research notes into a structured brief
---
# Summarize Research
Read all files in the research/ folder and create a structured
summary with key findings, notable quotes, open questions, and
recommended next steps.

Save the summary to research/summary.md

That’s it. No programming. No server process. Just plain old text.

Skills files are stored in Markdown format (.md), a file format like .docx or .txt.

It’s a file format designed with human readability in mind.

When you create a file with the .md extension and the contents above, and drop it in the right place, Claude will use these instructions next time you type “summarize my research” in Claude Code.

Skills work really well because of a couple of reasons:

1. They’re just text. Anyone who can write a clear set of instructions can create a Skill. They’re the standard operating procedures (SOPs) your AI agents will run their workflows on.

2. They’re cheap on context. Claude only loads Skill descriptions (the first two lines in the example above) when you start working with Claude. The full instructions load on demand, when you actually invoke the Skill.

3. They’re portable. Skills in Claude Code follow the Agent Skills open standard, which means the same Skill file works across Claude Code, GitHub Copilot, and any other tool that adopts the standard. Write once, use everywhere. There are already 200+ community-built skills available, and Anthropic maintains an official collection covering document creation, brand guidelines, and more.

4. They’re easy to iterate on. Since Skills are plain text, updating them is as fast as editing a document. I tweak mine constantly based on outcomes—tighten an instruction here, add an edge case there. The feedback loop is immediate.

They do have a couple of limitations:

• They can’t reach outside your machine. Skills have no way to connect to external APIs, databases, or cloud services on their own. They work with whatever is available in your local environment—files, shell commands, and whatever tools Claude already has access to.

• They’re inherently non-deterministic. Because Skills are natural language, Claude will interpret and execute them slightly differently each time. For creative workflows this is a feature. For processes that need to produce identical output every time, it’s a liability.

• Auto-invocation is unreliable. Several practitioners have documented that Claude’s automatic skill activation works roughly 50% of the time—“basically a coin flip.” You can mitigate most of this through clear naming and fine-tuning descriptions, but it can be a friction point communicating with your agent.

When you run into any of these, you might want to take a look at MCP servers instead.

MCP servers: connecting your local AI workspace to the world

Where Skills teach Claude how to work, MCP (Model Context Protocol) servers give Claude access to the data, software, and routines your business runs on.

An MCP server is a small software program that exposes external tools and data sources to your AI agents through a standardized protocol.

I wrote about the potential of MCP when Anthropic first launched it—calling it “the app store for AI agents.”

That prediction has largely played out: there are now hundreds of pre-built MCP servers for services like GitHub, Slack, Notion, PostgreSQL, and many more.

MCP servers are the software shell around your AI agents—their outer layer.

They give your AI—

• External connectivity. MCP is a great way to give Claude secure access to APIs, databases, and cloud services. Need to query your CRM? Search your company’s Notion workspace? Pull data from a PostgreSQL database? Use an MCP server.

• Deterministic tool execution. Unlike Skills, MCP tools contain written code. This means that when Claude calls an MCP tool, the execution is programmatic—the same input produces the same output every time.

• Agent-agnostic portability. Like Skills, MCP is an open protocol. A Postgres MCP server will works with Claude Code, Cursor, ChatGPT, and any other MCP client. You’re not locked into a single vendor’s ecosystem.

• First-party support. Many SaaS companies now ship official MCP servers. Notion, GitHub, Atlassian, Sentry, and others maintain their own—meaning you get supported, up-to-date integrations rather than community hacks.

However, MCP servers also come with several downsides:

• Setup complexity. First-time MCP configuration can easily take an hour: managing dependencies, setting environment variables, debugging permission issues. On Windows with WSL, the experience is significantly worse. Even experienced developers find the initial setup friction real.

• Maintenance burden. Every MCP server is another piece of software to maintain, update, and secure. API changes, dependency updates, authentication token rotations—it can add up.

• Security vulnerabilities in third-party servers. Community servers are a mixed bag—and in the worst case are built specifically to harvest your data or credentials. In general, I recommend either using first-party (official) MCP servers, or having Claude Code generate your own MCP servers.

Skills vs MCP—a simple decision framework

Use Skills when:

• The workflow is entirely local (reading/writing files, text processing, research synthesis)

• You need Claude to follow a specific procedure or methodology

• You want something you can create and iterate on in minutes

• The workflow involves creative judgment or open-ended analysis

Use MCP servers when:

• You need to connect to external systems (CRM, databases, APIs)

• Login credentials are required

• You need repeatable tool execution

• A first-party MCP server already exists for the service

• The integration is core to your workflow (worth the setup investment)

Use both when:

• You need external data and a specific procedure for processing it

This last case is where it gets interesting.

The real power: combining the two technologies

The developers who’ve published the most thoughtful analysis all arrive at the same conclusion: Skills and MCP are complementary, not competing technologies.

MCP provides the what. Skills provide the how.

Here’s a concrete example from my own setup.

Today I built a cold outreach system for my AI agency using Claude Code.

Total additional cost: €7 a month (a second Google Workspace subscription for a dedicated email address) + €29 one-off for Mailreach to warm up that dedicated inbox and ensure deliverability.

I’m using Attio CRM’s free plan, which is more than sufficient for my use case.

Costs will increase mainly depending on how much Apollo.io credits I need for company and contact data—note: I’d be paying for their data, not their software.

The MCP layer (connectivity):

• An Apollo.io MCP server lets Claude search for and enrich company and contact data using Apollo’s ~300M people lead database.

• An Attio MCP server (first-party, maintained by Attio) connects Claude to my CRM: it updates company and contact entries, deal stages, adds research notes, writes draft emails and checks pipeline status.

• A desk-research MCP server connects Claude to Perplexity for additional company research on their AI strategy and services—which is directly related to the services I offer through my AI agency.

The Skill layer (procedure):

• A /lead-source Skill tells Claude exactly how to run the outreach workflow: research the company first, determine who to contact based on company size and service angle, get their verified email, draft a 2-3 sentence message following specific rules, and save everything to a structured file and to Attio.

Without the MCP servers, the Skill has no ingredients to work with.

Without the Skill, the MCP servers are just raw tools with no workflow logic.

Together, I type /lead-source Philips products and Claude:

1. Checks if research already exists for Philips

2. Gets company data from Apollo (headcount, ARR)—converts USD to EUR

3. Runs desk research via Perplexity if needed

4. Searches Apollo for CEO, CTO, VP Engineering, Head of AI contacts

5. Enriches the top 1-2 contacts with verified emails

6. Drafts personalized messages referencing specific company details

7. Saves everything to a structured outreach file

8. Creates company and contact entries in Attio

The emails themselves are sent out manually through Attio after a final review and edit, using the dedicated email account for cold outreach.

Each email is reviewed manually to ensure quality and give a personal touch if needed.

And because I send these emails through Attio, I can track company contact moments and schedule follow-up sequences if I think they are a good fit for my services.

All of this saves me roughly 20-30 minutes of manual work per contact—while keeping the outreach itself personal, human and real.

The 4 hours I spent figuring out the right workflow and setting up the email account, Attio, the Skill and the MCP servers allowed me to email 7 companies in just 15m—instead of the 3-4 hours it would have taken me to do this manually.

Given that I want to reach out to 5-10 companies every morning, the time investment in setting this up will have paid itself back tomorrow morning.

The n=1 software argument

Why did I bring all this up in light of the “SaaSpocalypse”?

Because of what those plugins represented: AI agents that perform infinite variations of knowledge work when given the right Skills and the right connections.

The SaaS companies that lost the most market value last week built their businesses on a specific model—per-seat licensing for specialized tools that do one thing well.

CRM. Legal review. Financial modeling. Marketing automation.

As you could see from the example above, an AI agent with the right Skills and MCP connections can not only do the same things—it can do them better.

Without the need to resort to a clumsy SaaS drag-and-drop workflow builder.

Without needing to learn to work with yet another opinionated, poorly executed SaaS interface that tried to appease to everyone and ends up working for no one.

There are some limitations and caveats of course—I am using Attio and Mailreach for some of the things that are harder to replicate (or just not worth the time and effort).

But for most systems of action—for actually getting things done—an AI agent with the right Skills and MCP servers will beat 9 out of 10 SaaS software products.

Want to build workflows like this for your own business—but not sure where to start? AI operators is my 4-week 1-on-1 coaching program where we set up your AI workspace together. You tell me what eats your time—client research, content creation, outreach, admin—and we build the automations together. No technical background needed. Most clients get back 5-10 hours per week. DM me on Substack to see if it’s a fit.

---

Last Week in AI

• The S&P 500 software and services index fell nearly 9% over five trading sessions, with $285 billion in market cap vanishing in just 48 hours. Salesforce shed 14% in five trading days—roughly $30 billion—and is now down 26% year-to-date. SAP plunged 16%. Thomson Reuters fell 18% in a single session. ServiceNow dropped 11% despite forecasting subscription revenue above Wall Street estimates, and is down 28% year-to-date. Oracle suffered its worst eight-day stretch since 2002, falling 15.8% in one week. Atlassian slid 12.6%, LegalZoom cratered 20%, and HubSpot is down 39% year-to-date. Even Indian IT services firms like TCS (-7%) and Infosys got caught in the downdraft. Markets have dubbed it the “SaaSpocalypse”, and while Gartner analysts argue the reaction is premature, the underlying logic is hard to dismiss: why pay per-user fees for specialized software when an AI agent can do the same work at a fraction of the cost? Anthropic, meanwhile, is raising at a valuation north of $350 billion.

• On the other end of the same bet, the same earnings week revealed the staggering scale of Big Tech’s AI wager. Amazon, Alphabet, Meta, and Microsoft collectively disclosed roughly $650 billion in planned 2026 capital expenditures, a 60% increase year-over-year with the vast majority going to AI chips, servers, and data centers. Amazon alone committed $200 billion ($50 billion above Wall Street estimates) and saw its stock drop 8% in response. CEO Andy Jassy’s defense: spending too little on AI would be a far greater mistake than overspending. Markets are not yet convinced, with the four companies losing over $950 billion in combined market value during the week. In total, nearly $2 trillion vanished in a single week—SaaS investors fear AI works too well, Big Tech investors fear it doesn’t work well enough to justify the spend.

• OpenAI’s platform play ambitions were bolstered by a model, a platform and a partnership release last week: GPT-5.3-Codex is the coding engine—25% faster than its predecessor, top scores on SWE-Bench Pro and OSWorld (64.7%), and the first model to hit “High” on OpenAI’s cybersecurity preparedness framework. It is also the first model that meaningfully built itself, with the Codex team using early versions to debug training, scale GPU clusters, and diagnose evaluations. Frontier is the control plane—an enterprise platform where companies build agents via natural language, connect CRM systems and data warehouses into a unified semantic layer, and monitor performance through admin dashboards tracking success rates and audit logs. HP, Oracle, State Farm, and Uber are early customers. And a $200 million multi-year partnership with Snowflake embeds OpenAI models natively into Cortex AI across all three major clouds, giving 12,600 customers access to build agents over governed enterprise data.

• In other news: Anthropic shipped Claude Opus 4.6 with a 1-million-token context window, multi-agent coordination, and dramatically improved knowledge-work benchmarks. They also released a Powerpoint integration for Claude everybody’s raving about. Mistral shipped Voxtral Transcribe 2, a pair of speech-to-text models hitting state-of-the-art accuracy at $0.003/minute—the real-time variant is open-weights under Apache 2.0 and designed for on-device deployment.