FinWikiPublic-source finance field guide
日EN

Claude Code extension component architecture

ConfidenceCertain
Updated2026-05-26
Review by2026-10-30
Sources3
Machine-translatedOriginal (JA)
#agent-economy
On this page

Wiki route

This entry sits under AI Agent payment protocols overview · a 7-protocol survey. Read it with ERC-7715 overview · Wallet Permissions and AI Agent auto-payment for adjacent context and AP2 · Google Agent Payments Protocol overview for the broader system boundary.

[!info] TL;DR Source: 5 architecture diagrams received from 2026-04-16 (the WeChat official account “童童的 AI 创业圈”). The following is the full textualization of the image content.

1. The 6 extension components

Claude Code’s Agentic Loop (Plan → Act → Observe → Reflect) acquires capabilities through 6 types of extension components.

Component Role Use case Example
CLAUDE.md Holds persistent context, auto-loaded every session Project Specs (project conventions, global rules) Use pnpm; test before commit
Skills Reusable capability modules (knowledge, instructions, workflows), loaded on demand Reference Docs / Standard Tasks “/deploy to run deployment”
Subagent An independent sub-Agent. Runs in an isolated context and returns a result summary Context Isolation / Parallel Tasks Parse a file and return a conclusion
Agent Teams Multiple Agents collaborate, sharing a task list and communicating with each other Parallel Research / Complex Dev Launch multiple reviewers
MCP Connects to external systems and tools, accessing external data/actions External Data / Actions DB query; Slack send
Hook An automation script outside the loop. Triggered on specific events Deterministic Automation Run ESLint after changes

2. Distinctions between components

Skills vs Subagent

Aspect Skills Subagent
Essence Reusable knowledge or flow Independent execution unit
Core value Reuse content across different tasks Context isolation
Fitting scenario Reference material, standard flows Large-scale code analysis or parallel tasks

CLAUDE.md vs Skills

Aspect CLAUDE.md Skills
Loading method Auto-loaded every session Loaded on demand
File reference Supports @path Supports @path
Can trigger a flow No Yes
Best use Global rules Reusable flows

MCP vs Skill

Aspect MCP Skills
Essence External-system connection protocol Knowledge or flow
Capability provided Tool interface and data access Usage and business logic
Example Slack, DB Code-review flow

They are usually combined: MCP provides the connection, Skills define the usage.

3. Plugins & Marketplace

Components can be packaged as a Plugin for distribution.

Plugin
  ├── Skills
  ├── Hooks
  ├── Subagents
  └── MCP Servers
         ↓
    Marketplace (package and distribute extension features)

Skills use namespaces (e.g. /my-plugin:review) to avoid collisions.

4. Component hierarchy and merge rules

Components stack in 4 layers (additive):

User level
  ↓
Project level
  ↓
Plugin level
  ↓
Managed policy level

The CLAUDE.md files of all layers are loaded and merged into the conversation context. Conflict resolution: more specific rules take precedence.

Combination usage examples

Combination Behavior Example
Skills + MCP MCP provides the connection, Skills define the usage Skills describe the DB schema
Skills + Subagent Skills launch multiple Subagents /audit checks security and performance simultaneously
CLAUDE.md + Skills Global rules + on-demand knowledge CLAUDE.md defines the API conventions
Hook + MCP Hook calls an external system Notify Slack after a critical file change

5. Context cost

Each extension component consumes the model’s context window.

Component Load timing Loaded content Context cost
CLAUDE.md At session start All content Every request (high)
Skills At session start + on invocation Description + all content Relatively low
MCP Servers At session start Tool definitions Every request (high)
Subagents At creation Independent context Isolated from the main session
Hooks On trigger None by default Zero

A balance between capability and cost is required. Tips: set disable-model-invocation: true and the Skill is loaded only on manual invocation, with zero cost.

Potential connection to the agent payment layer (under consideration)

The following are early-stage associations, subjects for future consideration.

  1. Provide payment capability as an MCP Server: any Agent calls Agent Payment Infrastructure Research Report via the MCP protocol (check_balance, authorize_payment, get_price, dispute). The lower layer depends on ERC-4337 Overview · Application-Layer Implementation of Account Abstraction
  2. Monetizing the Marketplace: after becoming Skill marketplace monetization, it can be used as the backend for billing/payment (for the choice of payment currency, see USD Stablecoin Interchange Market)
  3. Tiered budget management: user level → project level → Agent level → task level → single invocation, mapped to the component hierarchy
  4. Financial cost tracking: similar to context-cost tracking, track the financial cost of each Skill invocation (how much was spent, whether it is worth it)

These directions are unconfirmed, recorded only.

Textualization compiled · 2026-04-16

Discovery

Keep reading

Related

Read next

Links here