Skip to content

Agent Skills vs MCP — different layers, complementary by design

A practitioner walk-through (Shaw Talebi, narrated as a Dutch-language summary) that resolves the common misread of Anthropic's Skills launch as a competitor to MCP. The video lands on the now-canonical answer: Skills supply the what to do, MCP supplies the how to connect.

Video

Source: https://www.youtube.com/watch?v=6wdvSH61xGw

Transcript notes

Key beats from the talk:

  • MCP framed as the "USB-C of AI" — a universal connector that standardises how a model reaches external tools, data, and prompt templates. Caveat: each connected tool's schema occupies context whether the tool is used or not.
  • Skills framed as instruction management — progressively-disclosed folders that load only metadata up front and pull body/resources on demand, so token cost stays near zero until a skill is triggered.
  • Three-level loading is the load-bearing design choice for Skills: name + description always in context (~100 tokens per skill); body loads on description match; bundled files load on explicit reference.
  • They compose — a Skill that, when triggered, invokes MCP-exposed tools is the working stack for mid-2026 agents.
  • The "Skills replace MCP" framing was a launch-week reaction from Simon Willison, tied specifically to MCP's tool-schema token bloat; Willison himself noted Skills require a filesystem and code-execution sandbox to work, i.e. Skills sit a layer above MCP.

Synthesis

The Model Context Protocol (MCP) is an open standard introduced by Anthropic on 25 November 2024 to give any AI application a uniform way to read data, call tools, and fetch prompt templates from external systems. The protocol uses a client-server architecture over JSON-RPC 2.0, with two transports — stdio for local processes and Streamable HTTP for remote servers — and three server-side primitives: tools (actions the model can invoke), resources (data the model can read), and prompts (reusable templates). On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI and backed by AWS, Google, Microsoft, Cloudflare and Bloomberg. The donation announcement reported more than 10,000 active public MCP servers, 97 million monthly SDK downloads, and first-class adoption in ChatGPT, Cursor, Gemini, Microsoft Copilot, VS Code, Sourcegraph Cody, Zed, and Claude.

Agent Skills were introduced by Anthropic on 16 October 2025 as a different abstraction — not a competitor to MCP but a complement. A Skill is a directory containing a SKILL.md file (YAML frontmatter, required name and description, optional Markdown body) plus optional supplementary markdown, scripts, and resources. The mechanism is progressive disclosure across three levels: at session start only the name and description of every installed skill enter the system prompt; the body of SKILL.md loads only when the model judges the skill relevant to the active task; bundled files and scripts load on demand through the agent's filesystem and code-execution tools. Anthropic published Skills as an open standard on 18 December 2025, making it portable across hosts that meet the filesystem + code-execution requirement.

The "competitor or complementary" question crystallised when Simon Willison published "Claude Skills are awesome, maybe a bigger deal than MCP" on launch day, arguing Skills are dramatically more token-efficient than MCP — a few dozen tokens of metadata per skill versus tens of thousands of tokens of tool schemas that an MCP server can dump into context. Willison himself flagged the dependency that resolves the apparent conflict: Skills only work where the agent has a filesystem and a code-execution sandbox, which makes them a layer above MCP rather than a substitute for it. Subsequent practitioner write-ups have converged on the same answer. Skills supply procedural knowledge — what to do and in what order — while MCP supplies the wire to external systems — how to connect to a Jira instance, a SharePoint tenant, or a Stripe account. The pragmatic stack is a Skill that, when triggered, calls MCP-exposed tools to get its work done.

Anthropic addressed MCP's context-cost problem directly in late 2025 and early 2026 by shipping Tool Search and Programmatic Tool Calling, allowing MCP server tools to be discovered progressively rather than loaded up front into the system prompt. That narrows the token gap with Skills and removes much of the original heat from the "Skills make MCP obsolete" framing, while leaving the layering intact. The two primitives now coexist in mainstream agent runtimes — Claude Code, Claude.ai, the Messages API, GitHub Copilot in VS Code — and Anthropic's own positioning of Skills as cross-platform packaging of organisational expertise treats MCP as a peer infrastructure layer rather than an obstacle.

The "USB-C of AI" framing repeated across vendor blogs and analyst posts does not appear in Anthropic's own donation or launch copy; it is community shorthand for MCP's role as a universal connector. Wikipedia notes a more conservative technical analogy — MCP "has been likened to OpenAPI" — and that comparison ages better as the protocol grows past simple wiring into governance, multi-tenant isolation, and audit. The pragmatic implication for a developer audience is that the USB-C metaphor is fine for pitches but starts breaking down once you ask hard questions about OAuth flows, tenant boundaries, or prompt-injection risk surface.

Two misconceptions are worth flagging because the video addresses both. First, "Skills replace MCP" is false: Skills do not connect to third-party SaaS, do not manage OAuth or tenant isolation, and assume the host already has the tools to invoke. Second, "MCP makes Skills obsolete" is also false — even with progressive tool discovery shipped in late 2025, MCP does not package procedural workflows, code, or reference assets the way SKILL.md does. Skills package institutional know-how (a runbook for a specific deployment, a brand-guidelines authoring routine, a particular PDF extraction strategy) in a portable folder; MCP packages access to systems. They solve different problems.

Governance and security considerations skew differently per layer. MCP carries the heavier multi-tenant, OAuth, audit, and prompt-injection risk surface — Wikipedia's MCP article tracks known prompt-injection and lookalike-tool risks documented since April 2025. Skills surface supply-chain risk instead, because a skill folder ships executable code and reference files that a host will load on triggering. An arXiv survey from February 2026 (Xu & Yan; single-source, treat as plausible-but-unverified) reports that 26.1% of community-contributed skills contain vulnerabilities and proposes a tiered trust-and-lifecycle framework. The mitigation patterns are different per layer too: pinning and signing for Skills; OAuth scoping and per-tenant key isolation for MCP.

The honest current view for an architect in mid-2026 is to write Skills that call MCP tools — not choose one over the other. The working model in Anthropic's own documentation and across the practitioner community is a two-layer agent stack: MCP for connectivity to external tools, data, and prompts under vendor-neutral governance via the Linux Foundation, and Skills for portable, progressively-disclosed procedural expertise that lives on the host filesystem and invokes those tools when triggered. The "USB-C of AI" line still serves as a marketing analogy but is increasingly stretched as both primitives mature.

Changelog

  • 2026-05-10 — Scaffold created from https://www.youtube.com/watch?v=6wdvSH61xGw (Type C, awaiting transcript)
  • 2026-05-11 — Synthesised from transcript + deep research; confidence pending → 93 (Type C transcript + 5 Type A primary sources + 2 Type B and 1 Type C corroborating sources); related_pages updated from anthropic-agent-skills (tombstoned) to claude-agents, agents-overview