AI 原生开放教材

外观

Claude Code Ultimate Guide工程手册·英文原文 ↗原件 ↓仓库 ↗

11.3 When to Build vs Use ​

For most developers, Claude Code CLI provides the right balance of power and simplicity. But understanding when to use pre-built agents versus building custom agents helps you choose the right tool for your needs.

Pre-Built Agents ​

Claude Code, Cursor, Windsurf, Goose: Ready-to-use CLI or GUI tools optimized for coding workflows. Install and start working immediately.

Use when:

  • You need a coding assistant now
  • Standard workflows (implementation, debugging, refactoring)
  • Want managed updates and community support
  • Prefer fixed costs (subscriptions) over usage-based pricing

Agent Builder Frameworks ​

Google ADK, LangChain, Vercel AI SDK: Code-first toolkits for building custom agents from scratch.

Key features:

  • Multi-LLM support: Use Claude, GPT, Gemini, or local models (often via unified APIs)
  • MCP-compatible: Consume the same MCP servers as Claude Code (3,000+ available)
  • Custom workflows: Design agent behavior for domain-specific tasks
  • Embedable: Integrate agents into your applications

Use when:

  • Building a product around AI agents (not just using one)
  • Need multi-LLM fallback strategies (e.g., Claude for code, GPT for other tasks)
  • Require custom execution flows beyond coding assistance
  • Want full control over prompts, memory, and orchestration
FrameworkFocusClaude SupportMaturity

Direct API ​

Anthropic API, OpenAI API: Raw LLM access without agent scaffolding.

Use when:

  • Building simple chatbots or single-purpose bots
  • Maximum flexibility, minimal dependencies
  • You're handling tool calls, memory, and orchestration yourself

Decision Tree ​

Need coding assistant? → Claude Code
Building agent product? → ADK/LangChain
Simple chatbot only? → Direct API
Multi-LLM required? → Framework (ADK/LangChain)
Custom workflow logic? → Framework
Standard dev tasks? → Pre-built agent

MCP: The Common Standard ​

All frameworks (Claude Code, ADK, LangChain) support the Model Context Protocol (MCP). This means:

  • MCP servers you build work across tools
  • 3,000+ MCP servers available ecosystem-wide
  • Tools, prompts, and resources are portable

Example: A mcp-server-github works with Claude Code, ADK-built agents, and LangChain agents identically.

Resources: