Map review: Agent Skills vs MCP — challenge an edge

[alsoleg89]

Live artifact

• Interactive map: https://alsoleg89.github.io/doubt/examples/agent-skills-vs-mcp.html
• Editable evidence record: https://github.com/alsoleg89/doubt/blob/main/examples/agent-skills-vs-mcp.doubt.json
• Content receipt: 0444afe44c523678f2ad8eb7267e0d7c7a392709921abf16474e68d2ef5a3991

Current position

Default to an Agent Skill for reusable procedure and bundled context. Add MCP for live or dynamic tools and resources; combine them when the Skill should orchestrate an MCP-backed capability.

The map is grounded in the Agent Skills specification, GitHub's client documentation, the Agent Skills discovery RFC, and the current MCP architecture and security guidance. Its source snapshot is dated 2026-07-31.

What to challenge

Please challenge a specific claim or reasoning edge, especially one of these:

1. Skill boundary: reusable procedure and on-demand supporting material fit an Agent Skill.
2. MCP boundary: live tools, dynamic resources, and prompts fit a client/server protocol.
3. Combined pattern: a Skill can teach a workflow that calls MCP-backed capabilities.
4. Counterpressure: remote MCP adds authorization and operational risk; a Skill may also execute scripts and is not automatically safe.

Useful review format

Edge or claim:
Verdict: supports / contradicts / qualifies / missing
Source URL:
Exact section, line, or paragraph:
Why the current map overstates or misses the evidence:

A broad “Skills are better” or “MCP is better” preference is not enough to change the map. A bounded source region or reproducible cross-client behavior is.

The largest explicit unknown is still open: there is no committed cross-client benchmark here showing that one non-trivial Skill and one MCP integration behave consistently across major coding agents.