Claude Cowork vs Claude Code: Which One for Which Job

Use Claude Code when the result of the work is a change to code: a fix, a feature, a commit or a pull request. Use Claude Cowork when the result is anything else a knowledge worker hands over: a report, a spreadsheet, a slide deck, a sorted folder or a summary of your inbox. Both take an outcome rather than a single prompt, plan the steps and work through them, and Anthropic describes Cowork as bringing Claude Code’s agentic capabilities to work beyond coding. The difference is who they are built for, where they run and what they can reach.

Who each one is for

Claude Code is for people who work in a code repository. Anthropic’s Claude Code overview↗ (opens in a new tab) calls it an agentic coding tool that reads your codebase, edits files, runs commands and works with your development tools. It assumes you know what a diff, a branch and a test run are, even if you have never opened a terminal; if that is you, start with Claude Code for beginners.

Cowork is for everyone else on the team, and for developers doing non-coding work. You describe an outcome, step away and come back to finished work, with no terminal involved. A finance lead reconciling two exports, an operations manager building a weekly pack, or a founder sorting a year of receipts is the audience.

Where each one runs

• Claude Code runs in a terminal, in VS Code and JetBrains IDEs, in the Code tab of the Claude desktop app, and on the web at claude.ai/code, where sessions run in the cloud and keep going after you disconnect. The terminal, IDE and desktop sessions run on your own machine. How the desktop, terminal and VS Code versions differ is covered in Claude Code desktop vs CLI vs VS Code.
• Cowork runs in the Claude apps. Anthropic’s page on Cowork on web, desktop and mobile↗ (opens in a new tab) lists desktop, web, mobile and the Chrome side panel. Its sessions run on Anthropic’s servers, in beta, so work continues in the background and scheduled tasks run with no device online.
• Cowork reaches your computer only through the Claude desktop app, which has to be open for local file access, browser use and computer use.

Cowork is also changing shape. Anthropic has announced that Cowork and chat are becoming one Claude↗ (opens in a new tab), rolling out to Pro and Max first, after which everything Cowork does is available from any conversation and there is no separate mode to pick. Claude Code stays a product of its own.

What each one can touch

• Files. Claude Code works in the project folder you open, and can read and edit anything in it. Cowork reads and writes local files through the desktop app without manual uploads, and works with files saved to your Claude account everywhere else.
• Commands. Claude Code runs shell commands on your machine, or in a cloud VM for web sessions: tests, builds, git. Cowork runs code and shell commands in an isolated environment on Anthropic’s servers, not on your computer.
• Other apps and services. Both use MCP. Claude Code connects to MCP servers you configure per project or per user. Cowork uses Claude’s connectors, skills and plugins, chosen from the plus menu in the message box.
• The browser. Cowork on desktop uses a browser built into the Claude desktop app by default, or Claude in Chrome when it should work in your own signed-in browser. Claude Code can drive Chrome through the same extension, mostly for testing a web app it has just built.
• Your screen. Computer use, where Claude operates other desktop apps, is in beta in both Cowork and Claude Code in the desktop app, on Pro and Max plans only.

Permissions and approvals

Both let you choose how often Claude stops to ask, and both keep a few stops you cannot turn off.

• Claude Code has permission modes↗ (opens in a new tab): Manual (default), which asks before most edits, commands and network access; acceptEdits; plan, which blocks edits until you approve a plan; auto, where a classifier model reviews each action instead of you; dontAsk for locked-down scripts; and bypassPermissions, for isolated containers only. From v2.1.283, auto is the starting mode for terminal and VS Code sessions. Shift+Tab cycles modes in the terminal.
• Cowork has three, listed on Anthropic’s page on using Cowork safely↗ (opens in a new tab): Manually approve, Automatically approve, where Claude checks each action for safety before it runs, and Skip all approvals. It always asks before permanently deleting files, and asks before using each app on your computer.
• Both can be governed centrally on Team and Enterprise plans: Claude Code through managed settings, and Cowork through an organisation-wide switch and approval-mode controls.

The practical rule is the same for both: start cautious for anything that sends messages, spends money or deletes, and loosen only for work you can undo. In Claude Code, git is the undo button. In Cowork, keep originals and point it at copies until you trust the task.

What each one hands back

Claude Code hands back changed files in a repository, and usually a commit or a pull request with a message it wrote. You review a diff. Cowork hands back finished documents: spreadsheets with working formulas, presentations, formatted documents, a renamed and sorted folder, or a summary. You review a file. Anthropic notes that Cowork sessions cannot be shared with others, so the file is the thing you pass on.

Plans, by name only

Both are on Pro, Max, Team and Enterprise. Neither is on the free plan. Claude Code can also run on an Anthropic Console account billed by usage, and the terminal and IDE versions support third-party cloud providers. On Enterprise, Cowork on web and mobile depends on the admin enabling it. Individual features differ by plan, so check before you buy for one: computer use, for instance, is Pro and Max only today.

Five jobs, sorted

1. Fix a failing test and open a pull request. Claude Code. The job lives in a repository, needs the test suite run, and ends in a diff someone reviews.
2. Turn three CSV exports into a reconciled spreadsheet with a summary tab. Cowork. The output is a spreadsheet with formulas, and nobody needs to see a terminal.
3. Write release notes from last week’s merged pull requests. Claude Code if the notes live in the repository as a changelog file; Cowork if they end as a document or an email to customers.
4. Every Monday, collect figures from connected tools into a status pack. Cowork, with a scheduled task. It runs in the cloud with the computer asleep and hands back the same file each week.
5. Check that a new sign-up page works, then fix what breaks. Claude Code with Claude in Chrome: build, open the page, find the fault and fix it in one session.

Many jobs split cleanly into both halves. A customer complaint becomes a bug that Claude Code fixes, and a reply and an account note that Cowork drafts. The split is by output, not by person. For more tasks on the Cowork side, with prompts to copy, see Claude Cowork examples.

Using both on one team

The trap with two agents is two private histories. Claude Code’s sessions live on one developer’s machine, and Cowork’s sessions belong to one person’s account and cannot be shared. Nobody else can see what either was asked to do, what it finished, or what it left half done. The fix is to keep the work list outside both, on a board each can reach over MCP.

fenbs is a small task board built for that. Tasks sit in four lanes, To Do, Next Up, In Progress and Completed, and are a feature, an enhancement or a bug. Claude connects as a custom connector, and Claude Code with one command; either signs in through the browser and holds your role on the board, narrowed by the scopes you tick. Every change is recorded in History under the assistant’s own name, so a task Cowork commented on and one Claude Code moved are told apart from each other and from a person. A good shared rule: the agent comments what it produced on the task and fills in how it was tested, and a person moves it to Completed.

claude mcp add --transport http fenbs https://fenbs.ai/api/mcp