Concepts Overview¶
Core building blocks that explain how Claude Code works and how to configure it.
Documents¶
| Document | Summary |
|---|---|
| Context window | How Claude manages active information, compaction, and context size by model |
| Project | What a Claude Code project is, memory hierarchy, team collaboration |
| MCP servers | Connecting external tools and data sources via the Model Context Protocol |
| Tools | Built-in tools, the agentic loop, permissions, and settings.json reference |
| Models | Model aliases, switching, effort levels, extended context, prompt caching |
How they relate¶
Model ──────────────────► sets reasoning depth + context window size
│
▼
Context window ◄──────── Tools feed results back into it each step
│
▼
Project (working dir) ◄── MCP servers extend what tools can reach
- Model determines what Claude can reason about and how far back it can see.
- Context window is the live workspace — tools write results into it, compaction trims it.
- Project defines the directory scope and loads the memory layer on top.
- MCP servers expand the tool surface beyond built-ins.
- Tools are the actions Claude takes inside the agentic loop.
For persistent behavior configuration (memory, hooks, agents, skills) see Persistence.