The starter guide tells you to install Claude Code and do a tiny project. That is still the best version of this if you can. But plenty of people cannot, and they usually assume it closes the whole thing off. It does not.
This page is for you if:
- You do not own a personal computer, and your only screens are a phone or a tablet.
- You have a work laptop you are not allowed to install software on, which is most work laptops.
- You have a computer, but it is old, slow, or shared, and you would rather not fight it.
The assistant does not have to run on your computer. It can run on a rented one, and you drive it from a browser tab.
What actually changes
In the normal setup, the assistant runs on your machine and works on a folder of your files. Here it runs on a small computer in the cloud that gets created when you start a task and thrown away when you are done. Same loop: you describe, it acts, you review.
The difference is that the files need somewhere to live between sessions, because the rented computer does not stick around. That somewhere is GitHub, and it is the one new idea on this page.
A folder becomes a repo
A GitHub account is free, and a “repository” is just a folder that lives on the internet and remembers every version of itself. You do not need to learn git to use one. You make it in a browser in about a minute, and from then on it is the folder your assistant works in.
Yes, this is an extra hurdle. It is also the starter guide's central idea arriving early instead of late: the files are the brain, the AI is just the hands. On this path you have to set up the brain first. That is not the worst way to learn it.
Get one thing right at the start: make the repository private. Public is an option on the creation screen and you do not want your notes indexed by search engines.
The two ways to do it
Both work from a browser on any device. Both need a paid subscription. That is the real cost of this path: there is no free tier that does the whole loop.
- Claude Code on the web. Go to claude.ai/code, connect your GitHub account, describe a task. It works in an isolated cloud machine, pushes a branch, and shows you the changes to review. Sessions keep running if you close the tab. It is also a full surface in the Claude iOS and Android apps, under the Code tab, where you can pick a repository and start a task, not just watch one. Pro, Max, and Team plans. Cloning repositories and opening pull requests requires GitHub specifically.
- Cursor cloud agents. The same shape, at cursor.com/agents on any device, plus a native iOS app and an Android web app. Needs a paid Cursor plan. Works with GitHub, GitLab, Bitbucket, and Azure DevOps if you have a reason to prefer one of those.
I would start with Claude Code on the web, for the same reason the starter guide picks Claude Code: fewer moving parts while you are learning the loop. Cursor is the better answer if you already pay for it, or if your code is not on GitHub.
It can reach your calendar and your email
This surprises people, so it is worth saying plainly. Cloud sessions can use connectors, which are the integrations you turn on at claude.ai/customize/connectors: Google Calendar, Gmail, Drive, Slack, Linear, and a growing list of others. Connector traffic goes through Anthropic's servers rather than the cloud machine's own restricted internet connection, so it works without any network configuration on your part. You choose which connectors a session gets.
Then there are routines, which are the part worth knowing about. A routine is a saved prompt plus a repository plus a set of connectors, running on a schedule. It runs in the cloud whether or not any device of yours is on. You set them up at claude.ai/code/routines, on Pro, Max, and Team. The minimum interval is one hour and there is a daily cap on runs.
A concrete version, because a friend asked for exactly this: keep a file in your repo listing your recurring payments and their due dates. A routine runs every morning, reads that file, cross-checks your calendar, and tells you what is due in the next few days. That is a genuinely useful thing to own, and none of it needs a computer.
Two cautions, and I mean these. Routines run unattended, with no approval prompts, so the prompt has to be specific about what success looks like. And every connector left attached is one the assistant can write with, not only read from. Trim the list to what the job needs. Best habit: have it write its findings into a file first and read them yourself for a couple of weeks. An automation that can only ever be wrong in a document is a very different thing from one that can be wrong in someone else's inbox.
What is worse about this path
- Your files are a click further away. You can browse them on github.com or in GitHub's mobile app, so you are not flying blind. It is just slower than opening a folder, and you build a feel for the project more slowly as a result.
- Review happens in a diff. Instead of watching edits land in files you can see, you read a summary of what changed in red and green. A fine way to review, a slightly worse way to learn, because it shows you the change rather than the thing.
- Sessions are more transactional. You send a task, it goes away, it comes back. That pushes you toward bigger and vaguer requests, which is the wrong instinct early on. Keep tasks small even when the interface makes small feel inefficient.
- It costs money from day one. The local path does eventually too, but you can poke at it for free first. Here you cannot.
None of that changes what you learn. Describing clearly and reviewing honestly are the skills either way, and they transfer intact to a laptop the day you have one.
Doing this from a phone
A phone is a supervision console, not a workstation. It is very good for checking on something that is running, reading what it did, making a small correction, or answering a question the assistant stopped on. It is miserable for composing the first careful description of a new project, and you will write a worse prompt than you are capable of.
If a phone is all you have, write that description in a notes app first and paste it in. That one habit is most of the difference.
What to do this weekend
- Make a free GitHub account.
- Create one private repository. Call it something like
notes. Let it add a README file when it offers. - Subscribe to one of the two tools above and connect it to your GitHub account.
- Pick one tiny project from the starter guide's list. The organizing ones fit this path best, since they are mostly text files. Ask for one small piece of it, not the whole thing.
- Read what it did before you accept it. Every time. That is the skill.
Leave connectors and routines until after that. They are the reason this gets addictive, but they make much more sense once you have done the loop once by hand.
If you get stuck on any of this, write to me at me@viyatjhaveri.com. The setup steps are where people quietly give up, and they are also the part I can unstick in two minutes.
— Viyat