← Back to recent questions

The short version is that they are two different tools wearing the same family name. Claude Code is a tool for people who write software. Claude Cowork is a tool for everyone else. They are both "agents", which means software that does a multi-step job for you rather than just answering a question, but they live in different places and are built for different people.

For the job you described, reading email as it arrives and drafting replies, the answer is Cowork. It connects to Gmail, it can run on a schedule, and it can write a draft reply and then wait for you to approve it before anything is sent. You do not need to write a line of code to set that up. One caveat to get out of the way early: Gmail is the email service Cowork connects to. If your mail lives in iCloud or Outlook, this route needs a rethink, and that is worth checking before you spend a cent.

So what actually is each one?

Claude Code runs in a "terminal", the plain text window programmers use to type commands to their computer. Point it at a folder of code and it can read the whole project, write and change code, run the tests that check the code works, and fix what it breaks, while you approve the changes as it goes. It is genuinely good at this and sits near the top of the industry's standard coding benchmark. If you write software for a living, it is worth trying. If you do not, the terminal alone will feel like being handed a plane's cockpit when you only wanted to cross the road.

Cowork is the one this site has written about before, on the AI agents page. It runs as an ordinary desktop app on Mac or Windows. You point it at a folder and give it a plain-English brief, and it reads your files, creates documents, reorganises things, and works through multi-step tasks while you get on with something else. On a paid plan it also connects to outside services like Gmail. The clearest way to hold the difference: regular Claude tells you how to do something, Cowork does it, and Claude Code does it but only for code.

One honest note before going further. Cowork is not free. It comes with the Claude Pro plan, around A$32 a month, not the free tier, and Claude Code is bundled with the same paid plans. Whichever way you go, this is a paid project.

Which one for automating email, and why?

Two routes. If you want your inbox triaged and replies drafted without becoming a part-time programmer, that is Cowork. If you are a developer who wants a custom pipeline running on a server, untouched by whether your laptop is awake, that is Claude Code or Anthropic's developer toolkit, and you will be writing and maintaining real software to get there.

For almost everyone asking this, the honest recommendation is the first route, and not just for ease. The email job is mostly about judgement, not engineering. The hard part is never "can the computer read my inbox". It is "does the draft make sense, and would I be embarrassed if it went out". Cowork keeps you in that seat by default. Claude Code will happily let you build something that removes you from it, which is the wrong thing to remove from email.

How would the email automation actually work?

At the simplest level, you connect Gmail to Cowork, then save a "scheduled task", which is just a saved instruction that runs on a timer, say every hour during the working day. The instruction might read: "Look for new emails that need a reply and draft a response to each in my tone, based on the thread. Do not send anything. Save the drafts for me to review." Cowork wakes up on the timer, reads the new mail, and produces the drafts.

Where the drafts land is your choice. The most cautious option is to have Cowork save them as files in a folder, so nothing touches your real inbox until you copy it across yourself. Either way the loop is the same: it drafts, you read, you decide.

One limitation to know up front. Cowork runs on your own machine, so the computer has to be awake and the app running for a scheduled task to fire. If the laptop is shut when the task is due, that run is skipped. This is the practical reason a developer with heavier needs eventually reaches for a server-based route, which does not care whether your laptop is open. For a personal inbox the awake-laptop limit is usually fine. For a business process that must run at 3am, it is not.

How do I keep it from sending something stupid?

This matters more than the setup, and it is where you should spend your attention. Start with the one setting that does most of the work. By default, Cowork treats sending an email as an action that needs your approval: it drafts the reply and shows it to you rather than sending. You can switch it to send automatically, but for email you should not, at least not until you have watched it work for a good while. Leaving that setting where it starts is the single most important safety decision in the project.

A draft can be wrong even when it reads well. The most common failure is that the model misreads the thread and answers the question it expected rather than the one actually asked. The fix is to read the original email, not just the draft, before you approve. Confident is not the same as correct, and the polish hides the mistake.

The harder cases are ambiguity and contradiction. When an email is unclear, or an earlier message in the thread disagrees with a later one, the assistant will usually pick one reading and write as if it were certain, unless you have told it not to. So a useful line in your instruction is: "If an email is unclear or the thread contradicts itself, do not draft a confident answer. Summarise the conflict and leave it for me." That turns a silent guess into a visible question.

Finally, privacy. Connecting an AI assistant to your inbox means it can read your inbox, and your mail is some of the most sensitive material you own. On a paid plan the data protections are stronger, but the content is still processed on Claude's servers. If your email holds client confidences, health information, legal matters, or anything you are obliged to keep private, think hard before you connect the whole account, and consider limiting the automation to a single label rather than the lot.

What I would avoid

I would avoid switching the send setting to automatic on day one, or in the first month: the whole value of the draft-first design is that a human still reads the email before it leaves. I would avoid pointing it at your entire inbox before testing it on one label first. I would avoid reaching for Claude Code or a custom pipeline just because it sounds more serious; for a personal inbox that adds maintenance and risk without adding much you will feel. And I would avoid trusting a draft because it is well written. Fluent and right are different things, and email is exactly where that gap costs you.

A simple test before you trust it

Before you rely on this, run it against email you can already judge. Take ten messages you have already answered, let the assistant draft replies, and compare them with what you actually sent. Then feed it three deliberately awkward ones: an email that is vague about what it wants, a thread where an earlier message disagrees with a later one, and a complaint dressed up as a question. If the easy ten come back close to what you would write, and the awkward three come back as "this is unclear, here is the conflict" rather than a confident wrong answer, it is working. If it sails past them with breezy certainty, it is not ready, and the approval step stays in place.

The verdict

Claude Code is the coding tool. Cowork is the everyday desktop agent. For reading email and drafting replies, use Cowork: connect Gmail, write a scheduled task that drafts but never sends, and leave the approval setting exactly where it starts. Try it first on a single label and a handful of emails you have already answered. The danger is not the technology failing loudly. It is a fluent, confident, slightly wrong reply going out because you stopped reading. Keep yourself in the loop, and this is one of the genuinely useful things an agent can do for you today.