LLM - Claude Skill vs Agent vs MCP

Posted May 20, 2026

By Grace L

22 min read

Claude Skill vs Agent vs MCP

Claude Skill vs Agent vs MCP

Overview 概述

在 Claude Code 生态中，有三种核心概念用于扩展 Claude 的能力，它们处于不同的层次，服务于不同的目的。

In the Claude Code ecosystem, three core concepts extend Claude’s capabilities. Each operates at a different layer and serves a distinct purpose.

MCP：给 Claude 一双手 — 连接外部系统，提供新能力 MCP: gives Claude hands — connects to external systems, provides new capabilities
Skill 技能：给 Claude 一本操作手册 — 告诉它如何高效完成特定工作流 Skill: gives Claude an operating manual — tells it how to execute a workflow consistently
Agent 代理：给 Claude 一支团队 — 并行运行多个独立的 Claude 子实例完成复杂任务 Agent: gives Claude a team — runs multiple independent Claude sub-instances in parallel for complex tasks

概念定义 Concept Definitions

Skill 技能

技能是一个存放在 .claude/skills/ 目录下的 Markdown 文件（SKILL.md），通过 YAML 元数据中的 description 字段自动触发。技能在主 Claude 实例的同一上下文窗口中运行，提供工作流知识，而不增加新的系统能力。

A Skill is a Markdown file (SKILL.md) stored in .claude/skills/, auto-triggered by matching the description field against the user’s request. It runs inside the main Claude instance’s context window and provides workflow knowledge without adding new system capabilities.

触发方式 Triggering: description 字段关键词匹配 / description field keyword match
运行位置 Runs in: 主 Claude 上下文 / main Claude context window
编写方式 Authored in: Markdown，无需编程 / Markdown, no coding required
目的 Purpose: 让 Claude 的某类工作流始终一致、专业 / make a class of workflows consistent and expert

注意 Note: 技能目录里可以有 agents/ 子文件夹，但那只是存放子工作流指令文件的命名惯例，不是真正的进程。 A skill directory may contain an agents/ subfolder, but that is only a naming convention for sub-workflow instruction files — not real subprocesses.

Agent 代理

代理（Subagent）是由主 Claude 通过 Agent 工具显式生成的一个全新的、独立的 Claude 子实例。每个 Agent 拥有自己的上下文窗口、工具访问权限，独立执行任务后将结果返回给编排它的主 Claude。多个 Agent 可以并行运行，这是处理大规模并发任务的关键能力。

An Agent (Subagent) is a brand-new, isolated Claude sub-instance explicitly spawned by the main Claude using the Agent tool. Each Agent has its own context window and tool access, executes a task independently, then returns a result to the orchestrating Claude. Multiple Agents can run in parallel — this is the key capability for large-scale concurrent work.

触发方式 Triggering: 主 Claude 显式调用 Agent 工具 / orchestrating Claude explicitly calls Agent tool
运行位置 Runs in: 独立上下文窗口（新 Claude 实例）/ isolated context window (new Claude instance)
编写方式 Authored in: 主 Claude 或技能中的编排逻辑 / orchestration logic in the main Claude or in a skill
目的 Purpose: 并行执行、任务隔离、分而治之 / parallel execution, task isolation, divide-and-conquer

MCP (Model Context Protocol)

MCP 是一个基于 JSON-RPC 2.0 的开放协议，运行为独立的服务器进程。它向 Claude 暴露三种原语：工具（可调用函数）、资源（只读数据源）、提示模板。MCP 赋予 Claude 它本身不具备的外部能力，例如查询数据库、调用第三方 API、读写文件系统。

MCP is an open protocol (JSON-RPC 2.0) running as a separate server process. It exposes three primitives to Claude: Tools (callable functions), Resources (read-only data sources), and Prompts (pre-built templates). MCP gives Claude external capabilities it does not have natively — querying databases, calling third-party APIs, reading/writing file systems.

触发方式 Triggering: Claude 的工具调用机制 / Claude’s tool-use mechanism
运行位置 Runs in: 独立服务器进程 / separate server process
编写方式 Authored in: Python、Node.js、Go 等 / Python, Node.js, Go, etc.
目的 Purpose: 连接 Claude 到外部系统 / connect Claude to external systems

三层对比表 Three-Way Comparison

维度 Dimension	Skill 技能	Agent 代理	MCP
所处层级 Layer	提示 / 指令层 Prompt / instruction	执行 / 并发层 Execution / concurrency	基础设施 / 连接层 Infrastructure / connectivity
提供什么 Provides	工作流知识 Workflow knowledge	并行执行能力 Parallel execution	新能力（工具、数据）New capabilities
运行在哪 Runs in	主上下文 Main context	独立子上下文 Isolated sub-context	独立服务器进程 Server process
上下文共享 Context	共享主上下文 Shares main context	全新隔离上下文 Fresh isolated context	在 Claude 上下文之外 External to context
并行能力 Parallelism	否 No	是 Yes	N/A
编写方式 Authored	Markdown	编排逻辑 Orchestration logic	编程语言 Programming language
是否需要编程 Coding	否 No	取决于编排 Depends	是 Yes
独立工作 Works alone	是 Yes	是 Yes	是 Yes
触发机制 Triggering	description 字段 Field match	显式调用 Explicit call	工具调用机制 Tool-use

用户场景 User Scenarios

场景 1 — 只用 Skill：统一工作流 Scenario 1 — Skill only: standardize a workflow

情境 Situation: 每次写 PR 描述，格式都不一样，团队审查效率低。

Every PR description looks different, slowing down code review.

解决方案 Solution: 创建 pr-writer 技能。每次说 “帮我写 PR 描述”，Claude 自动按团队模板（概述 → 变更类型 → 测试方式）生成规范内容。

Create a pr-writer skill. Every time the team says “write a PR description,” Claude automatically follows the template (overview → change type → test plan).

为什么不用 Agent / MCP Why not Agent/MCP: 任务是单一的、顺序的，不需要并行，也不需要连接外部系统。

The task is single and sequential — no parallelism needed, no external system to connect.

场景 2 — 只用 Agent：大规模并行分析 Scenario 2 — Agent only: large-scale parallel analysis

情境 Situation: 一个代码仓库有 50 个文件，需要逐一检查是否有 SQL 注入漏洞。如果顺序执行，需要等很久。

A repo has 50 files, each needing SQL injection checks. Doing them sequentially takes too long.

解决方案 Solution: 主 Claude 把 50 个文件分成 10 组，同时生成 10 个子 Agent，每个 Agent 检查 5 个文件，最后主 Claude 汇总所有 Agent 的结果。

The main Claude splits 50 files into 10 groups, spawns 10 subagents simultaneously — each checks 5 files — then the main Claude aggregates all results.

为什么不用 Skill / MCP Why not Skill/MCP: Skill 在主上下文中顺序运行，无法并行。MCP 提供工具，但不解决并发问题。

A skill runs sequentially in the main context — it cannot parallelize. MCP provides tools but doesn’t solve concurrency.

场景 3 — 只用 MCP：连接外部系统 Scenario 3 — MCP only: connect to an external system

情境 Situation: 想让 Claude 能直接查询 Postgres 数据库，拉取报表数据。

The team wants Claude to query a Postgres database directly and pull report data.

解决方案 Solution: 搭建 Postgres MCP 服务器，暴露 query_db 工具。Claude 调用该工具执行 SQL，无需人工复制粘贴数据。

Set up a Postgres MCP server exposing a query_db tool. Claude calls the tool to run SQL — no manual copy-paste needed.

为什么不用 Skill / Agent Why not Skill/Agent: Skill 和 Agent 都没有内置数据库连接能力，MCP 是唯一能赋予这种能力的层。

Neither skills nor agents have built-in database connectivity. MCP is the only layer that can grant this capability.

场景 4 — Skill + Agent：专家编排并行子任务 Scenario 4 — Skill + Agent: expert orchestration of parallel subtasks

情境 Situation: 安全审查工作流需要：读取每个模块的代码 → 按 OWASP 清单分析 → 汇总报告。模块很多，单线程太慢。

A security review workflow needs to: read each module → analyze against OWASP checklist → produce a consolidated report. Many modules, sequential is too slow.

解决方案 Solution: 创建 security-reviewer 技能，技能的工作流中编排多个 Agent：每个 Agent 负责一个模块，按技能内嵌的 OWASP 清单审查，完成后主 Claude 汇总所有发现。

Create a security-reviewer skill. The skill’s workflow orchestrates multiple Agents: each Agent handles one module against the embedded OWASP checklist, then the main Claude consolidates all findings.

security-reviewer Skill
        ↓
为每个模块生成一个 Agent / Spawn one Agent per module
        ↓  ↓  ↓  ↓  ↓  (并行 parallel)
  [Agent A] [Agent B] [Agent C] [Agent D] [Agent E]
  模块1分析  模块2分析  模块3分析  模块4分析  模块5分析
        ↓
  主 Claude 汇总，生成最终安全报告
  Main Claude consolidates → final security report

场景 5 — Skill + MCP：专业工作流使用外部工具 Scenario 5 — Skill + MCP: expert workflow using external tools

情境 Situation: 已经有 GitHub MCP（可以调用 create_pr、add_comment），但 Claude 每次用这些工具的方式不一致，评论质量参差不齐。

GitHub MCP already exists (create_pr, add_comment tools). But Claude uses those tools inconsistently, producing variable-quality comments.

解决方案 Solution: 创建 pr-review 技能，内嵌标准化工作流：拉取 diff → 分类问题严重级别 → 按模板用 add_comment 写结构化评论。MCP 提供工具，Skill 提供使用这些工具的判断逻辑。

Create a pr-review skill with a standardized workflow: pull the diff → classify issue severity → write structured comments via add_comment following a template. MCP provides the tools; the Skill provides the judgment for using them.

场景 6 — Agent + MCP：子代理并行调用外部系统 Scenario 6 — Agent + MCP: subagents calling external systems in parallel

情境 Situation: 需要同时查询 5 个不同地区的数据库，每个数据库在不同的 MCP 服务器上，合并成一张全球报表。

5 regional databases (each on a different MCP server) need to be queried simultaneously and merged into a global report.

解决方案 Solution: 主 Claude 生成 5 个子 Agent，每个 Agent 连接对应地区的 MCP 服务器，查询后返回结果。主 Claude 合并 5 个结果。

The main Claude spawns 5 subagents — each connects to its region’s MCP server, runs the query, returns the result. Main Claude merges all 5 results.

场景 7 — 三者结合：全自动端到端安全扫描 Scenario 7 — All three: fully automated end-to-end security scan

情境 Situation: 一个完整的仓库安全扫描：需要从 GitHub 拉代码、并行分析各模块、按标准工作流生成报告。

A full repo security scan: pull code from GitHub, analyze modules in parallel, generate a report following a standard workflow.

解决方案 Solution:

security-auto Skill (技能编排整个流程 / skill orchestrates the entire flow)
        ↓
为每个模块生成子 Agent / Spawn subagent per module
        ↓       ↓       ↓
  [Agent A]  [Agent B]  [Agent C]   (并行 parallel)
      ↓           ↓          ↓
  GitHub MCP  GitHub MCP  GitHub MCP  → 拉取模块代码 / pull module code
  SAST MCP    SAST MCP    SAST MCP    → 运行静态扫描 / run static analysis
      ↓           ↓          ↓
   返回发现 Return findings
        ↓
  主 Claude 汇总 / Main Claude consolidates
        ↓
  GitHub MCP create_pr → 提交安全报告 PR / submit security report PR

技能提供标准化工作流；Agent 提供并行执行；MCP 提供 GitHub 和扫描工具的访问权限。

The Skill provides the standardized workflow. Agents provide parallel execution. MCP provides access to GitHub and the SAST scanner.

Agent vs 多个 Terminal Session Agent vs Multiple Terminal Sessions

这是一个很常见的认知误区：多开几个 Claude Code 终端窗口，不就和用 Agent 一样了吗？

A common misconception: opening multiple Claude Code terminal sessions does the same thing as using Agents. The conceptual similarity is real — but the operational difference is significant.

两者都是多个独立的 Claude 实例，各自有独立的上下文窗口 Both involve multiple independent Claude instances, each with its own context window
两者都可以同时处理不同的任务 Both can handle different tasks simultaneously

核心差距：谁来协调 Core Difference: Who Coordinates

	多个 Terminal Session	Agent 子代理
协调者 Coordinator	你（人工）You (manual)	主 Claude（程序化）Orchestrating Claude (programmatic)
规模上限 Scale limit	~5 个（人的注意力有限）~5 (human attention is the bottleneck)	50+ 个（程序无此限制）50+ (no programmatic limit)
结果汇总 Result aggregation	手动复制粘贴 Manual copy-paste	自动回流给主 Claude Automatically returned to orchestrator
条件逻辑 Conditional logic	你看结果再决定下一步 You decide next step after reading results	主 Claude 自动分支、重试、组合 Orchestrator auto-branches, retries, combines
嵌入工作流 Embeddable in workflow	否 No	是（Skill 可编排 Agent）Yes (Skills can orchestrate Agents)
会话生命周期 Session lifecycle	持续存在直到手动关闭 Persists until manually closed	任务完成即结束（短暂）Ephemeral — ends when task is done
交互模式 Interaction mode	多轮对话 Multi-turn conversation	单次任务 → 返回结果 Single task → return result

更准确的类比 A Better Analogy

多个 terminal session = 你自己当项目经理，挨个打电话给几个外包，等对方回复，自己汇总报告。

Multiple terminal sessions = you acting as project manager — calling each contractor individually, waiting for replies, writing the summary report yourself.

Agent = 你雇了一个项目经理（主 Claude），它自动派活、追踪进度、收结果、写汇总，你只看最终输出。

Agents = you hired a project manager (the orchestrating Claude) — it assigns work, tracks progress, collects results, writes the summary. You only see the final output.

什么时候多开 Session 就够了 When Multiple Sessions Are Enough

如果满足以下所有条件，手动开多个 session 和用 Agent 的结果几乎一样：

If all of the following are true, multiple terminal sessions and Agents produce nearly identical results:

任务数量少（≤5 个）Task count is small (≤5)
任务之间相互独立，不需要结果汇总 Tasks are independent, no aggregation needed
不介意手动协调和复制结果 Manual coordination and copy-paste is acceptable
不需要条件逻辑（A 的结果决定是否执行 B）No conditional logic (A’s result determines whether to run B)

什么时候必须用 Agent When Agents Are Necessary

需要并行处理 10 个以上的独立子任务 10+ independent subtasks need to run in parallel
结果需要自动汇总后再做决策 Results must be automatically aggregated before the next decision
任务编排有条件分支（基于中间结果动态派发更多任务）Orchestration has conditional branching (spawn more tasks based on intermediate results)
工作流需要嵌入 Skill 中自动触发 Workflow needs to be embedded in a Skill and triggered automatically

需要连接外部系统（数据库、API、文件系统）？
Need to connect to an external system (DB, API, filesystem)?
  → 用 MCP

任务是否可以被分解成多个独立的并行子任务？
Can the task be broken into independent parallel subtasks?
  → 用 Agent（可以在 Skill 内编排）
     Use Agent (can be orchestrated within a Skill)

需要 Claude 每次都用同一套专业流程处理某类工作？
Need Claude to follow the same expert process for a class of work every time?
  → 用 Skill

以上都有？
All of the above?
  → Skill 编排 Agent，Agent 调用 MCP
     Skill orchestrates Agents, Agents call MCP

Key Takeaways

MCP 是连接层，赋予 Claude 外部能力。MCP is the connectivity layer — gives Claude external capabilities.
Skill 是指令层，嵌入领域专业知识，保证工作流一致。Skill is the instruction layer — embeds expertise and ensures workflow consistency.
Agent 是执行层，用于并行和隔离。Agent is the execution layer — used for parallelism and isolation.
技能的 agents/ 子文件夹只是命名惯例（存放子工作流 Markdown），不是真正的进程。The agents/ subfolder inside a skill is only a naming convention for sub-workflow Markdown files — not real processes.
三者互补，最强大的系统将三者结合：Skill 编排 → Agent 并行 → MCP 连接。All three are complementary — the most powerful systems combine all three: Skill orchestrates → Agent parallelizes → MCP connects.

References

_posts/51AI/CLAUDE/2023-04-24-Claude_skill.md — Claude Skill internals
_posts/51AI/AIAttack/2021-08-11-MCPSecurity.md — MCP security attack surface
Anthropic Claude Code Docs

51AI, CLAUDE

AI Claude Skill Agent MCP workflow multi-agent

This post is licensed under CC BY 4.0 by the author.