Why do we need MCP if skills exist now?

The simple answer is auth, but skills and MCP solve different problems. Skills package knowledge. MCP packages authenticated, scoped access on a user's behalf.

Before MCP, agents had three problems performing external tasks: integrations were bespoke (every client had its own tool interface), authentication meant handing over long-lived API keys, and agents couldn't discover what tools were available at runtime.

MCP standardized the interface and made OAuth a first-class path. With MCP, users can grant agents scoped access to their accounts—with revocation, auditability, and least-privilege tokens—without sharing API keys or credentials. Lee Robinson sums it up: "There's a lot of helpful data behind auth. MCP gives agents access to that data."

MCP's eventual ubiquity was far from obvious at the time of its release. Early critiques focused on security and adoption uncertainty. They then shifted to context bloat, which degraded agent performance as the number of tools provided increased. Despite this, MCP kept growing and today it is the standard most agent platforms converge on for third-party integrations.

Where MCP Falls Short

The context bloat issue never really went away.

Teams with large APIs ship massive MCP specs that dominated the context window. Anthropic notes that 50+ tools consume ~72K tokens before work begins. Dynamic tool loading (searching for the right tool at runtime) helps reduce this, but done naively it invalidates the KV cache. Manus calls KV-cache hit rate "the single most important metric for a production-stage AI agent." With Claude Sonnet, cached vs. uncached tokens show a 10x cost difference. Every time you dynamically swap tools in and out, you're potentially paying that 10x penalty.

But even if context bloat gets solved, there's a second problem: MCP does very little to help agents learn how to use tools well or perform tasks effectively. If you tell an agent to make you a component, it would build you one. But it'll import icons from a barrel file instead of using direct imports—loading 1,583 modules and adding 2.8 seconds to dev startup. Without a skill, agents default to what they've seen most in training data.

What Skills Solve

A skill is a markdown file that tells an agent how to perform a task effectively—this can include best practices, gotchas, decision trees, and tribal knowledge. Skills work extremely well because agents learned to code in filesystems. That's why every context mechanism has converged on files: CLAUDE.md, AGENTS.md, llms.txt, .cursorrules, etc. Skills extend this pattern to products and workflows.

Skills also by default don't have the cache-busting issues of MCPs—LLMs can intelligently load skills into their context as they need them. Agents are getting better and better at this, and researchers are getting SOTA performance out of LLMs just by getting them to manage their own context.

Among the most popular skills is Vercel's React best practices, which encodes dozens of rules across multiple categories including eliminating request waterfalls, reducing bundle sizes, and re-render optimization. Through what was likely a herculean effort, the Vercel team distilled years of production experience into a format agents can easily apply.

Are skills just documentation?

Yes. Skills are documentation organized and written for agents. Typically this knowledge is siloed in experts' heads, scattered across developer docs, or left undocumented entirely. With agents, the incentive to write it down is increased exponentially as they can instantly apply it.

The major problems with skills historically have been discovery and installation (I use historically loosely here, skills are still extremely new). Most skills are scattered across GitHub repositories, and until recently there was no standard way to discover and install them.

Vercel's skills CLI launched to solve these problems, gaining 75k npm downloads in its first week. npx skills add <owner/repo> works across Claude Code, Cursor, Windsurf, and others. You can now install a skill, keep it up to date with a CLI command, and search through popular skills.

On Mintlify, every documentation site now auto-generates a skill and hosts it at /.well-known/skills/. We also generate an MCP server at /mcp—giving agents direct access to your indexed docs without relying on web search. Skills can be managed with the skills CLI or agents can find and navigate them directly.

As skills become more standardized, distribution and discovery will improve—and it'll be easier to train models to use them correctly by rewarding successful tool use through reinforcement learning (RL).

The remaining gap is authentication. Skills do nothing to give access to authed data or perform actions on behalf of the user. This is what MCP is for—and skills can make agents dramatically better at using MCPs.

The Third Option: CLIs

There's a pattern emerging that combines the best of both approaches: authenticated CLIs, especially paired with skills.

Agents have been RL'd to be unreasonably good at using CLIs, and do so at the same level of proficiency as navigating filesystems.

CLIs handle authentication even more naturally than MCPs. OAuth flows, browser-based login, and credential storage are solved problems for CLIs. For example, PlanetScale's CLI authenticates with pscale auth login, which then opens up a browser and logs you in with one button click.

But auth isn't the only advantage. CLIs have been around for a long time, and as Mintlify Engineering Manager Nick Khami has pointed out, with CLIs you can expect standardized patterns around help menus, interactive prompts, and flags—while MCPs are still relatively new and vary wildly in quality.

CLIs are far from a silver bullet. They require more upfront work than MCPs, which have libraries like FastMCP that make spinning up a server relatively simple. And of course skills are even easier than MCPs, as they are just YAML files with markdown.

CLIs are better in terminal-based coding environments like Claude Code and OpenCode, but MCPs make more sense in GUI-based tools like Claude Desktop, where the users steering agents are likely to be non-technical.

What We Expect to Happen

MCP is here to stay. It's the standard for how agents authenticate and interact with external services. The spec will continue maturing, especially around context bloat.

Skills will grow exponentially. The prospect of millions of skills, each full of concentrated, actionable tribal knowledge, is one of the most exciting things happening in AI tooling right now.

The choice between skills and MCP is a false dichotomy. Skills, MCP, and CLIs are core tools for developers building for humans and agents.