Skip to content

Skills Over MCP

alex-krentsel's empirical finding from a month of openclaw hands-on: text-based skills beat MCP-server tools for most extensibility needs. "All this hype around MCP servers, adding more tools — really I think skills seem to be winning out."

The three-fidelity skill format

Anthropic-style skills have a staged structure that matches how the agent fetches context:

  1. Header (SKILL.md top, ~5 lines) — when the skill is applicable. Always loaded into the system prompt (up to 150 skills / 30k chars; filtered beyond that). Does not say how or what.
  2. Body (rest of SKILL.md, 10s–100s of lines) — what the skill does and how to do it. Fetched only if the agent decides the skill is relevant.
  3. Linked files — examples, assets, scripts. Fetched only after the agent has read the body and wants to execute.

This mirrors OS-level demand paging: the cheap header is always resident; the expensive body is paged in on reference; the executable linked files are paged in only at use.

Why skills beat MCP for most cases

  • Effectiveness. Skills are remarkably reliable in practice — the one-password skill, the YouTube-upload skill Ludwig discovered mid-task. They work.
  • Authorability. Non-technical users can write skills. They're just markdown. MCP servers require code, protocol conformance, a running service.
  • Self-discovery. Skill repos have ~46k GitHub stars already; the agent can scan them, propose installation, and add them autonomously (OpenClaw asks first by default). MCP servers are heavier to install and run.
  • Context-cheap. The header-only default means adding 100 skills costs almost nothing until one is used — MCP tools all claim slots in the tool list.

When you still want an MCP / tool

  • Work the LLM cannot do in pure text: talking to a GPU cluster, hitting a binary API with auth tokens, high-throughput I/O, deterministic structured output required.
  • Environment tooling Krentsel still prefers at the CLI layer: Google Workspace CLI, exc.dev CLI, Modal CLI, Claude Code via API key.

Broader implication

The bet is that as models get smarter, the cheapest extensibility wins, and the cheapest extensibility is "just write a markdown file that tells the agent when and how." This is a concrete case of design-over-implementation — formalism loses to prose because prose is what the agent speaks.

Pairs with

Cross-references