Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More

Owned by Aurel

AI narovinu bez zbytočného hypu

111 contributions to AI Automation Society
🚀New Video: How to Put Claude Code on a VPS in 10 Minutes
In this video I set up Claude Code on a Hostinger VPS so it runs 24/7 in the cloud, then connect it to Telegram so I can talk to my whole AI OS from my phone. I walk through the one-click Claude Code VPS, logging in with your subscription, installing bun, creating a Telegram bot, pairing it so only you can message it, and keeping it online after a reboot. AGENT SETUP GUIDE
0 likes • 43m
Worth one check after the setup. Everything your AI OS reaches on your laptop (the API keys, client folders and the Claude login) now sits on a rented server that's online all day, so the server's own login becomes the front door. I'd open /etc/ssh/sshd_config and turn PasswordAuthentication off once your SSH key works, and only copy over the folders you actually want to reach from your phone. The Telegram pairing covers who can message the bot, not who can log into the box.
✋ Where do you put the human in the loop?
Almost every serious build I do ends up with a point where a person has to say yes before anything continues. The hard part is not adding it. The hard part is putting it in the right place. Too early, and the human is approving something that has not been worked out yet, so they approve blindly. Too late, and they are asked to sign off on a finished result they can no longer really change. Where I have landed: the approval goes right before the first step that cannot be undone. Everything before it can be drafted, enriched, checked, rewritten as often as needed. The moment something leaves the building, a message sent, a record changed in a system someone else relies on, that is where the yes sits. The other thing I learned the hard way: make the approval fast. If saying yes takes three clicks and a login, people start approving without reading, and then the check is only theatre. One screen, the actual output, approve or reject. That is it. Where does the human sit in your builds, and do they actually read what they approve? 👇
0 likes • 1h
One thing to check behind that one screen is who can actually press yes. If the approve button sits in a shared Slack channel, anyone in the channel can click it, and in n8n the resume link a Wait step sends will continue the run for whoever opens it unless you turn on authentication for it. Then the log says approved, just not by the person you meant. I'd tie the button to the one approver and record the name from the click itself, not from who the request was sent to.
Google kills refresh tokens every 7 days in Testing mode
Google kills your refresh tokens every 7 days while your OAuth app is still in 'Testing'. I built my own command-line tool for Gmail, Drive and Sheets. gcloud couldn't do the job. Its token carries only Google Cloud scopes, and its OAuth client isn't even allowed to request Gmail or Drive scopes. I skipped a Claude connector on purpose, because it holds one account at a time and the whole point was covering more than one. So now my tool does the usual mail actions from list to trash, plus Drive find/download/upload and Sheets read/append. The part I care about most is what it can't do. Every outbound action (send, reply, draft-send, trash, sheets append) shows a preview and refuses to act without --yes. Plain drafts are exempt, a draft isn't outbound. It also can't permanently delete mail at all. The scopes are gmail.modify plus gmail.send instead of full mail access, so trash stays recoverable for 30 days and purge is impossible. I published the app to production to get rid of the 7 day expiry. Setup was finished on 2026-08-03 with both accounts verified on live data. Small gotcha: Drive's 'contains' search matches word prefixes, not substrings. --name "Guestlist" finds the file, --name "uestlist" misses it. My lesson is to put the safety limits into the permissions themselves, so I don't have to trust myself to be careful later. When you wire your own tools into email or Drive, do you lock them down with narrow OAuth scopes and confirm-before-send, or do you rely on being careful?
0 likes • 22h
Those scopes cover the write side well. On the read side gmail.modify still reads every message in both accounts, so whatever session calls the tool can pull any mail into its context, and from there it sits in that chat's history. If most of what the agent does is triage, a label filter inside the tool, where only messages under one label ever come back, puts the same kind of limit on reads that --yes puts on sends.
The Claude Code file that leaked secrets in 1,122 repos
Apiiro reported on Nov 5 that 1,122 public copies of one Claude Code file on GitHub hold a real secret. The file is the local settings file Claude Code keeps in your project's claude folder (settings local json). Every time you click "Yes, and don't ask again", Claude Code saves that command there. If the command had a key in it, the key got saved too. I opened mine. It was last written on August 11. It held a live Bearer token and a Google client secret, in plain text, both saved by "don't ask again" clicks I'd forgotten about. It never reached GitHub because one line in my gitignore file lists it. That one line was the whole gap between my repo and those 1,122. Claude Code 2.1.323 (Nov 12) fixed the cause. An approval that looks like it holds a secret now lasts for the session only. It doesn't clean up what's already in the file. If you use Claude Code, this takes five minutes: 1. Run git log --all on that file. Any output means it was committed at some point. 2. Untrack it with git rm --cached and add it to the repo's gitignore. 3. Rotate every key it ever held, even in a private repo. I'm rotating both of mine, even though the file was never pushed. A key that sat in plain text for three months had three months to be copied. When did you last open your own local settings file?
1 like • 22h
One more place to look while you're in there. The command that put the key into settings.local.json also ran inside a session, and Claude Code keeps every session as a jsonl transcript under ~/.claude/projects, tool calls included. That folder isn't in the repo, so no gitignore line covers it, and it rides along with every backup of your home folder. Rotating kills those copies too, but a grep for the old token there shows how far one approval spread.
I built a Windows app to run all my Claude Code sessions in one window - looking for honest feedback
Like a lot of you, I run several Claude Code sessions at once, and I kept losing track of which terminal was stuck waiting for me to approve a command. I'd come back ten minutes later and nothing had moved. So I built a Windows app for it, called Gozco. It runs all your Claude Code terminals in one window. The one that's waiting turns yellow, plays a sound and sends a Windows notification. Blue is working, green is done. It also shows tokens and cost per terminal, your 5-hour and weekly limit bars, and reads your reminders from an Obsidian note. It uses your own Claude subscription and never touches your global Claude settings, hooks or CLAUDE.md. No account, no telemetry: the usage numbers are read from Claude's own log files on your PC. The free version handles up to 4 terminals. There is a one-time Pro for phone notifications and approving from your phone, but the free one is the real tool. Honest question before anything else: do you even feel this pain, or do you manage several Claude Code sessions fine with just terminal tabs? I'd rather hear "I don't need this" than have you install it to be polite. And if you would use something like this, what is the one thing it must do? Tell me what's missing or what feels off. Site and download: https://jantecreative.github.io/gozcu-site/
1 like • 2d
The approve from phone part is the one I'd think hardest about. A notification has room for a line or two, and the command waiting for approval can be long with the risky bit at the end, a path to a .env file or a curl to somewhere. I'd make the phone screen show the full command and the folder it runs in before the approve button works, and keep the notification text itself short, since it passes through the push service and sits on the lock screen.
1-10 of 111
Aurel Babiš
4
8 points to level up
@aurel-babis-9673
Finance & Automation Engineer

Active 20h ago
Joined Dec 28, 2025
Powered by