// The guide

Claude Desktop app: chat, Cowork and Code in one window.

2026-07-15 · 5 min read · by Nabil BA-MOH

You open the Claude app and face three tabs — Chat, Cowork, Code — and no explanation of why the same AI needs three doors. So you use the one you know, ignore the other two, and quietly wonder what you're missing. Most people are running a three-engine machine on one engine.

Here's the guide Anthropic didn't write: what each tab is for, what each one can actually touch on your computer — that's the real difference — and what the desktop app gives you that claude.ai in a browser never will. It runs on macOS 12+, Windows 10+, and Linux in beta.

The mental model: three tabs, three levels of access

Forget features for a second. The three tabs are three trust levels for your files:

• Chat — sees nothing on your disk. It's claude.ai in a native window: you upload what you want it to read. Zero risk, zero reach.
• Cowork — sees a sandboxed copy of folders you explicitly connect. The agent works inside an isolated VM; your system stays outside the glass. (Full Cowork guide here.)
• Code — sees your project folder directly. No sandbox copy: real files, real edits, with permissions and checkpoints as the guardrails.

Pick the tab like you'd pick a collaborator: Chat to think, Cowork to delegate, Code to build. Once you hold that model, the whole app makes sense.

What desktop adds over the browser

Two things the web version structurally can't do:

• Local MCP servers — connectors that run on your machine and reach local tools and data, not just web services.
• Desktop Extensions (.mcpb) — those same local connectors as one-click installs, no config-file surgery.

Plus scheduling, session management, and an iOS companion: Dispatch a task from your phone, and the desktop or cloud executes it.

The Code tab: a full workshop, not a chat panel

If you only know Claude Code as a terminal command, the Code tab is the version with a cockpit:

• Parallel sessions via git worktrees — several tasks on the same project, each in its own isolated copy, side by side.
• Rearrangeable panes — chat, diff, browser preview, terminal (Ctrl+`), file editor, plan, tasks, subagents; build the layout your work needs.
• Diff review with line comments — read every proposed change like a pull request before it lands, and comment inline.
• PR/CI monitoring with auto-fix and auto-merge toggles.
• Environments — run local, remote, over SSH, or in WSL.

Shortcuts worth stealing: Cmd+/ lists them all; Cmd+N new session; Ctrl+Tab cycles sessions; Cmd+Shift+D opens the diff.

Not a developer? Doesn't matter — the same tab organizes files, drafts documents and automates the boring parts. That's a whole guide of its own.

Computer use: Claude takes the mouse (carefully)

The desktop app also ships computer use as a research preview on macOS and Windows: Claude sees the screen and operates apps directly. The design tells you Anthropic's threat model:

• Pro and Max only — deliberately not Team/Enterprise yet — and off by default, behind macOS Accessibility and Screen Recording permissions.
• Permission tiers per app class: browsers are view-only (real browsing goes through Claude in Chrome), terminals and IDEs are click-only, everything else can be fully driven.
• Per-app, per-session approval — you grant each app, each time (30 minutes for Dispatch tasks).
• Tool hierarchy: it prefers a connector over a dedicated tool over raw screen control. The mouse is the last resort, not the party trick.

That reluctance is the feature. An agent that grabs the mouse only when no cleaner path exists is one you can actually leave alone.

One window, one memory problem

FAQ