There is a new category of AI tool that goes well beyond a chat window. Instead of answering questions, these tools can do work on your computer: read and write files, run scripts, open a browser, and carry out multi-step tasks on your behalf. The leaders right now are OpenAI's Codex, Anthropic's Claude Code, and Anthropic's Claude Cowork. This is what each one is good at, what I would steer you away from, and how to tell whether any of them will actually help with what you do.
A note before the names: this category moves fast, and the specific rankings will shift. The part of this post that matters longest is the section on the three capabilities, because those are what separate a tool that helps from a tool that wastes your afternoon, regardless of which product is ahead this quarter.
The ones I recommend
Codex and Claude Code are both very good. They are the two I reach for, and either one will serve a non-technical person well. In my own use, Codex is a little friendlier for someone who does not want to think about the machinery, and Claude Code is a little better once you are working in earnest. That gap is small and personal, so do not agonize over it. If you are starting from zero, Codex is the gentler on-ramp. If you want the tool I personally get the most out of, it is Claude Code.
Claude Cowork is also good, and it is aimed at a slightly different person. Where Codex and Claude Code grew out of developer tooling and still carry some of that shape, Cowork is built more for general knowledge work. If the idea of a terminal makes you uneasy, it is worth a look.
The honest summary is that you will be well served by any of these three. The choice between them matters far less than understanding what they can do for you, which is the rest of this post.
The one I would steer you away from
When I tested Google's Antigravity, it behaved in ways that made it hard to recommend for this kind of work.
When I asked it to open a browser, it told me it could not, even though that capability exists. A tool that does not know what it is able to do is worse than a tool that simply lacks the feature, because you cannot trust its own account of itself.
When I asked it a simple question about whether it could read Word documents, it did not just answer. It went off and installed a pile of software packages and rearranged files on my system to chase the question, when a one-word answer would have done. That kind of overreach is exactly what you do not want from a tool that has access to your files. The risk is not that it fails a task. The risk is that it makes unrequested changes to your computer while trying.
Products improve, and this was my experience at the time I tested it, so it may get better. As it stood, I would not put it in front of someone who is not technical enough to notice and undo what it did.
The three capabilities that actually matter
Strip away the brand names and the real question is whether a tool can do three things. These are what determine whether an AI agent is genuinely useful for real work or just a fancier chat box.
It can open a real browser and let you log in. A surprising amount of useful work lives behind a login: a portal, an account, a web app that has no other way in. A tool that can open an actual browser window and let you authenticate can do those tasks. A tool that cannot is limited to whatever it can reach without signing in, which rules out most of the things you would actually want done. When you are evaluating a tool, this is the first thing to confirm.
It can run scripts when running a script is the right move. Sometimes the sensible way to do a job is a few lines of code: rename three hundred files, pull data from a spreadsheet, reformat a batch of documents. A capable agent can write and run that script, which is faster, more reliable, and far cheaper than having the AI grind through the work by hand one item at a time. Letting the tool run code for the mechanical parts is often the difference between a task that takes seconds and one that burns through your usage doing it the slow way.
You can extend it, and you know how. These tools become far more useful when you add to them. That means installing MCP services, which let the tool connect to outside systems and data, and adding plugins and skills, which teach it to do specific jobs well. There is also a distinction worth understanding between configuration that applies everywhere you work, called global, and configuration that applies to a single project. Getting this right is where a tool stops being generic and starts being yours. It is also the part most people never set up, which is why most people get a fraction of the value.
That last capability is the one that quietly separates people who get enormous value from these tools from people who find them a novelty. The tool out of the box is useful. The tool extended with the right connections and skills, configured for the work you actually do, is a different thing entirely.
What to actually do
If you just want to start, install Codex or Claude Code, and do not overthink which one. Use it for a real task, not a toy one, and notice where it saves you time. Once it has earned a place in your week, the next step is extending it, and that is usually where having someone set it up properly pays off, because the configuration and the skills are exactly the part that is fiddly to get right and easy to get wrong.