Chat, Cowork, and Code, Oh My!three Claude surfaces, three context windows, one model familyClaude Chat200K> upload reference docs> chat over knowledge> RAG behind the wall# auto-summarized# no operator leversCowork— not published —> grant local folders> shell in a sandbox VM> spreadsheets + slides# memory only inside# ProjectsClaude Code1,000,000> live working dir> real shell, real git> /compact /context# subagents own window# 5x the room to worksourced only from Anthropic's own docs
· 5 min read ·

Chat, Cowork, and Code, Oh My!

ai claude claude-code

A friend asked me this week how Claude Chat with Projects compares to Claude Code. I only use Claude Code, and I had not stopped to think the answer through. What I found, once I went to confirm everything from Anthropic’s own docs, was more interesting than expected.

Three surfaces, one model family, three very different ways of seeing your stuff. Every claim below links to its source on support.claude.com, code.claude.com, or claude.com/blog. Where the docs are silent, I say so.

Context Window

Bar chart comparing context window sizes per Claude surface: Claude Chat 200K, Claude Enterprise 500K, Claude Code 1 million, Cowork undocumented

Pro, Max, and Team share the same 200K window. Max buys more weekly usage, not more context. Enterprise gets 500K. Claude Code on Sonnet or Opus, opted in, sits at 1M. Cowork’s window is not published in any official article I could find.

What Each One Loads

Three columns showing what fills the context window for Claude Chat, Cowork, and Claude Code, plus how each surface behaves when the window fills

Knowing the size is not enough. What goes in matters more.

A Claude Project ships with custom instructions and an uploaded knowledge base of “documents, text files, or code snippets” (Projects help). When knowledge exceeds the window, RAG mode kicks in and a project knowledge search tool retrieves only the relevant chunks. Anthropic claims this expands capacity “by up to 10x.”

Claude Code enumerates exactly what fills its window: conversation history, file contents, command outputs, CLAUDE.md, the first 25KB of MEMORY.md, loaded skills, and system instructions. You also get explicit primitives, /compact and /context, plus subagents that run in their own fresh windows and return only a summary.

That is the operator gap. Claude Chat compresses invisibly while Claude Code shows you the budget.

Projects Are Not Folders

A Claude Project sounds like a project folder on disk. It is not.

What a Project actually contains, per the help:

  • Uploaded documents, text files, or code snippets
  • Custom instructions for every chat in the project
  • A grouped collection of chats

What a Project does not have, as far as official docs go: a GitHub connector wired into the knowledge base, a Google Drive sync, or any automatic mechanism for pointing at a live repo. None of the four help articles I read describe one.

Claude Code points at the working directory itself. Per the Claude Code docs, it has access to your live files, your terminal, your git state, CLAUDE.md, and auto memory. Your codebase is the knowledge. There is no upload step.

Translating across the two: custom instructions map roughly to CLAUDE.md, and project knowledge maps roughly to the file tree plus auto memory. What does not translate is git awareness. A Claude Project knows nothing about your branch, your last commit, or what changed since you uploaded that file. That is the gap a Code user feels first.

Cowork Is the Agentic Surface

Cowork is what my friend was actually trying to ask about. It runs only on the macOS or Windows desktop app, available on Pro, Max, Team, and Enterprise. Per Anthropic’s getting-started article:

  • Local file access in folders the user grants
  • Shell and code execution “inside an isolated virtual machine, separate from your main operating system”
  • Web search, spreadsheets, presentations, MCP connectors
  • “Memory is supported within projects but is not retained across standalone Cowork sessions”

Projects sit inside Cowork, the same way Projects organize chats on the web. There is no separate “Agent mode” toggle in the help docs.

One detail matters most for a Code user: Cowork’s shell runs in a sandboxed VM, not on your real machine. Claude Code runs in your real shell, with your real git tree. That is why Code can git push to your repo and Cowork cannot.

What Anthropic Has Not Published

Worth flagging, since these came up while researching:

  • A specific token-size for Cowork’s context window
  • A GitHub or Google Drive connector wired into a Project’s knowledge base
  • An exact RAG activation threshold inside a Project
  • Per-file or total-storage caps on Project knowledge

Each of those is genuinely undocumented as of today. If you find a definitive Anthropic source for any of them, send it.

Picking the Right Surface

How to pick:

  • Code work in a git repo. Claude Code is the only surface designed for that shape.
  • Curated reference documents with grouped chats over them. A Claude Project is built around exactly that pattern.
  • Desktop knowledge work mixing local files with spreadsheets and slides. Cowork, since the agent runs in a sandboxed VM that cannot touch your real filesystem.

I keep Claude Code as my default because everything I do lives in a working directory and I want the agent close to git. After this research, that has not changed.

For the friend who asked: this is not “which one is better.” It is which one matches the shape of your work.

Every claim above links to its Anthropic source. Where the docs are silent, I said so. Find me on Instagram or contact me.