Where Claude works, and how it reaches outside

Run claude in a folder and that folder becomes its world — the working directory. But the project isn't a cage. Claude reads memory from parent folders, loads your global ~/.claude, and can be handed extra roots — while anything genuinely outside the fence has to clear a permission prompt first. Step through it: the boundary is drawn, and every crossing is either pre-authorized or asked for out loud.

A nested-folder map of Claude Code's working directory: a green scope fence around my-app, a claude token, and reach lines out to ~/.claude and a blocked secrets/.env.

~/Users/micky

.claudeglobal · always in scope

CLAUDE.mdglobal memory

skills/

agents/

settings.json

projectsmonorepo root

CLAUDE.mdinherited memory

my-appcwd

CLAUDE.mdproject memory

src/

index.ts

utils.ts

package.json

shared-uisibling package

Button.tsx

secretsoutside the fence

.env

$claude

The working directory is the anchor

You ran claude inside my-app. That folder becomes the working directory — the root of everything Claude does. On launch it already reads my-app/CLAUDE.md for your project's instructions.

cwd~/projects/my-app

--add-dirnone

in scopemy-app

always~/.claude

Parent walk-up

Claude reads CLAUDE.md / AGENTS.md from the cwd and every parent directory up to your home folder. Project memory inherits from above.

Global ~/.claude

Global memory, skills, subagents, and settings in ~/.claude are loaded for every session, no matter which project you start in.

--add-dir

Add extra trusted roots at launch or mid-session. The boundary grows to include them, so reads and edits there happen with no prompt.

Permission prompt

Reach for any path outside the trusted roots and Claude must ask first. Allow once, allow always (added to settings), or deny — your call.

Where Claude works, and how it reaches outside

The fence, and four ways across it

The working directory is the anchor