I still prefer MCP over skills

A spirited debate brews over how AI agents should interact with the world, pitting the Model Context Protocol (MCP) against the burgeoning 'Skills' paradigm. The author advocates for MCP as a robust API abstraction for service integration, contrasting it with the perceived clunkiness and CLI dependencies of Skills. This discussion resonates deeply within the developer community, highlighting architectural preferences and practical challenges in building the next generation of AI tooling.

Score

Comments

Highest Rank

15h

on Front Page

First Seen

Apr 10, 3:00 AM

Last Seen

Apr 10, 7:00 PM

Rank Over Time

The Lowdown

The article "I Still Prefer MCP Over Skills" dives into the architectural choices for integrating Large Language Models (LLMs) with external services, presenting a strong case for the Model Context Protocol (MCP) over the increasingly popular "Skills" approach.

The author, a heavy AI user, expresses dissatisfaction with Skills, particularly when they necessitate the installation and management of Command Line Interfaces (CLIs). They argue that while pure knowledge-based Skills are valuable, relying on CLIs for service interaction introduces significant friction.

Advantages of MCP: The article highlights MCP's core philosophy as an API abstraction, allowing LLMs to interact with services without needing to understand the underlying mechanics. Its benefits include zero-install remote usage, seamless updates, robust authentication (e.g., OAuth), true portability across devices, inherent sandboxing, smart discovery of tools, and frictionless auto-updates.
Friction with Skills: The primary critique of Skills focuses on their reliance on CLIs. This creates deployment complexity, secret management nightmares, fragmented ecosystems, and context bloat (loading entire SKILL.md files). Many LLM environments (like standard ChatGPT or Claude web) cannot run CLIs, making such Skills "dead on arrival."
Proposed Ecosystem: The author suggests using MCP for connecting LLMs to services (e.g., Google Calendar, Notion, Xcode) due to its clean, strongly-typed interfaces and protocol-level handling of auth and updates. Skills, conversely, should be reserved for "pure" knowledge tasks, such as teaching LLMs how to use existing tools (like curl or git), standardizing workflows, or explaining secret management patterns.
Combined Approach: The ideal scenario, according to the author, involves using Skills as a "cheat sheet" or knowledge layer on top of MCP connectors. This allows LLMs to learn best practices and avoid common pitfalls when interacting with MCP-managed services, enhancing the overall experience.

The article concludes by emphasizing the need for standardized interfaces like MCP to achieve seamless AI integration, rather than a fragmented landscape of hacky CLIs.

The Gossip

Complementary Concepts: A Harmonious Handoff

Many commenters assert that MCP and Skills are not mutually exclusive but rather complementary, serving distinct purposes in the AI agent ecosystem. They suggest MCP handles the 'infrastructure' layer, providing standardized interfaces for external tools and data, while Skills manage 'orchestration' and high-level behavioral context. Some propose using both, with remote MCPs providing security and CLI-based Skills enabling programmatic execution, or even Skills explaining how to use MCPs. This perspective views the choice not as 'either/or' but as 'both/and,' depending on the specific problem being solved, such as local development vs. remote agent interaction.

Skills Supremacy: The CLI's Call

A significant portion of the discussion champions Skills, often arguing against MCP's necessity. Critics suggest MCP is merely a CLI wrapped in boxes, adding unnecessary complexity and overhead without significant security or simplicity gains. They highlight that modern CLIs can be secure (e.g., AWS CLI), and that skills can readily integrate with existing CLI tooling or even bundle code/scripts to call APIs directly, making a separate MCP layer redundant. Some believe the simplest solution (often CLI/Skills) will ultimately prevail due to Occam's Razor.

Practical Pitfalls: Reliability and Real-World Woes

Commenters share real-world frustrations with both MCP and Skills. Some report poor reliability with MCPs, experiencing timeouts or invalidated connections with desktop AI clients, especially when accessing remote servers. On the Skills side, agents are frequently observed ignoring clear instructions within Skills, even brief ones, leading to inefficiency. The issues of context bloat, secure secret management, and varying harness support for different MCP transport/auth mechanisms are also raised, indicating that neither approach is without its practical challenges in current implementations.

Defining Divisions: Where Do Capabilities Lie?

There's a lively debate around the actual capabilities of Skills, with some directly refuting the author's premise that Skills cannot call APIs or CLIs. Commenters clarify that Skills can indeed contain code, bundle scripts, teach agents how to use `curl` for API calls, or integrate with security features. This sub-discussion suggests a misunderstanding of Skills' full potential, indicating that many of the author's 'friction points' with Skills might be addressable through more advanced Skill implementations, potentially blurring the lines between the two approaches further.