There are four basic components:
Claude Code
Claude Code is Anthropic's agentic command-line tool for code generation, refactoring, and debugging. It runs where your code lives, and you configure its behavior through a CLAUDE.md file, custom skills, and MCP integrations.
The distinction that matters: this isn't a chat window you paste code into. It's an agent that reads your files, makes edits, runs commands, and checks its own work inside your project.
Where it shines:
Pair programming on a real codebase
CI/CD automation (think fixups, migrations, dependency bumps)
Code review bots that comment on diffs
Monorepo tooling where context spans many packages
If your task is "operate on this repository," Claude Code is the surface built for it. Agent SDK
The Agent SDK is a framework for building multi-agent applications. It hands you the primitives you'd otherwise write yourself: agentic loops, subagent orchestration, tool calling, and full lifecycle hooks.
The value here is that the hard parts of agent plumbing are already solved. You describe the agents and tools; the SDK runs the loop, routes between subagents, and gives you hooks to observe or intercept each step.
Where it shines:
Customer support agents that route, look up, and resolve
Research pipelines that fan out across sources and synthesize
Autonomous workflows that chain multiple decisions
If you're writing your own while loop around a model call and manually dispatching tools, the SDK is probably what you actually wanted. Claude API
The Claude API is the HTTP layer for programmatic access to the model. This is messages.create(), tool_use, structured JSON output, batch processing, and streaming. No agent framework, no CLI, just the model and your code.
This is the lowest-level, highest-control surface. You own the loop, the state, and the orchestration. In exchange, you get maximum flexibility and the ability to run at scale.
Where it shines:
Data extraction from unstructured text
Content generation pipelines
Classification at scale (batch processing is your friend here)
If you want a single well-defined transformation, run many times, with predictable structured output, the API is the direct path. MCP: Model Context Protocol
MCP is a standard, not a client. It defines how Claude connects to your databases, APIs, and services through tool and resource interfaces. You write (or install) an MCP server that exposes capabilities, and any MCP-aware client, including Claude Code and the Agent SDK, can use them.
The point of MCP is that you build the connector once and reuse it everywhere. Instead of wiring a database into three different apps, you write one MCP server and every surface can talk to it.