AIToday
Large Language ModelsOpen-Source AIDaily Dose of Data SciencePublished: Sep 18, 2026, 06:00 JST

MCP adds standard Agent Skills discovery and loading

MCP adds standard Agent Skills discovery and loading

3 Key Points

  1. What happened

    MCP now defines a standard way to discover and load Agent Skills from MCP servers, exposed as resources.

  2. Why it matters

    Agents can load only the skill needed for a task, cutting the context window pressure of upfront workflow instructions, with reusable workflows traveling with the server.

  3. What to watch

    The benefit hinges on whether clients read skills on demand rather than caching everything; the body names no rollout date or adoption figure.

WHO IT HITSTeams building AI agents on MCP servers can now ship workflow know-how alongside the tools they already expose, rather than maintaining separate docs, repos, or prompt files. The gain depends on client implementations that read skills on demand, the body suggests.

Not sure about something? Ask the AI

Summaries like this, in your inbox every morning.

Context & Analysis

The MCP addition sits on top of a connection layer that already standardized how agents reach tools, resources, and external systems. What was missing was a standard place to put the playbook: the reusable knowledge of how a workflow should be performed. Previously, that knowledge often lived separately in docs, repos, prompt files, or custom integrations, which meant it had to be maintained and versioned apart from the server that provided the capability.

Serving skills through MCP's Resources primitive changes that distribution model. Because SKILL.md, references, scripts, and examples are exposed as resources, an agent can first discover what skills exist, inspect their metadata, and then load only the one required for the current task. The article frames this as context window management: instead of loading every workflow instruction upfront, the agent pays only for what it uses, and reusable workflows travel with the server rather than with each client.

How much this matters in practice hinges on client behavior. On-demand loading only saves context if clients actually defer reading skills instead of prefetching them, and the body offers no adoption or rollout figures. For teams already running MCP servers, the practical question is whether their client of choice supports this discovery-and-load flow; if it does, workflow knowledge can be versioned and shipped alongside the tools themselves, and if it does not, the older split between capability and playbook is likely to persist.

FAQ
How are Agent Skills actually delivered through MCP?
They are served through MCP's existing Resources primitive. Files such as SKILL.md, references, scripts, and examples are exposed as resources the client can read on demand.
What is the difference between tools, resources, and skills in MCP?
The article's mental model is: tools are what the agent can do, resources are what the agent can access, and skills are how the agent should perform a reusable workflow.
What is the Nebius AI Builder Program mentioned alongside this?
It launched last week with $400+ in free credits for Nebius Token Factory, Tavily, Toloka, LangSmith, and more. It includes runnable cookbooks and training from top companies.
Daily Dose of Data ScienceRead Original Article

Get the latest Large Language Models news every morning

For example, today's edition would include:

  • OpenAI launches Astra for Law, a GPT-6 setup for legal researchSiliconANGLE AI · 2h ago
  • Google opens CC to families of six as shared AI agentSiliconANGLE AI · 2h ago
  • Anthropic shares pacing metrics, cites 30,000 agentsSiliconANGLE AI · 2h ago

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. Q&As are published on this page for other readers too.

Related Articles

Next articleOpenAI nears second Millennium Prize with Hodge conjecture