We've all been there. Close the lid, come back, everything's asleep. Agent mid-task, render stopped, SSH session dead. 🥸

Built Keepresso to control Mac sleep on your terms, dynamically.

✨ For programmers, vibe coders, AI agent users, render farms, headless Mac minis and Studios.

🆓 Free. Open-source. No license, no tracking. Just true brewing.

Keepresso.com

a while back I wired Claude Code to a Telegram bot so I could fix my server from my phone. then I kept adding one more thing. 🙂

today it's MyAgens: a self-hosted fleet of AI agents that lives on your machine. open source. 🫰

👑 you sit at the top, Atlas coordinates, and leads own their domains: devops, finance, research, anything you need. each lead can get its own Telegram bot. one message in, the whole fleet moves.

🖥 my daily use: "the website is down, fix it" sent from my phone. it reads the logs, finds the nginx typo, and asks before restarting. anything that changes the system waits for a tap.

🕯 the web panel adds live system graphs, a task board that delegates to agents, browsable memory, an encrypted vault, and a feed showing every diff as agents write code.

built on Claude Code. Grok CLI, Codex CLI, and local models work too (Ollama, LM Studio). ✏️

🔈 music: Blast by Aylex
ℹ️ source: github.com/gyorgysh/myagens

🎙 full story:
> gyorgy.sh/blog/myagens

install is one line, grab it at MyAgens.com

Keepresso more than a keep-awake app.

It fixes headless display issues, kills gaming lag, keeps your Slack status active, and a bunch of other smart triggers on top
☕️

💬Agentic coding is quietly becoming the new IDE.

A year ago AI was a chat window you copy-pasted from.

Now it reads the repo, writes, runs tests, fixes itself. 10x output, same hours. Still needs tuning for smarter indexes, local memory, references that push the model to refactor rather than pile on more bloat.

Lots of room to grow.

The role is shifting with it. Every senior engineer is becoming a project manager now, directing and training AI on their own repos. Less typing, more deciding what good looks like.

And it's becoming pay-to-win.
The smartest cloud models are a day/night difference over self-hosted or budget ones, and that gap decides what you can build.

As a solo builder or small team, the best model you can afford is your real ceiling.

The best model you can afford is your real ceiling.

Last post I said that, and I stand by it. The highest available models are a night-and-day difference over the budget ones, and that gap is real.

But which model you're on is only half of it. The other half is how you run it, and that's what decides whether you reach the ceiling or burn most of it on waste.

Three write-ups from this week on that second half:

1. The habits quietly running up your usage
Seven of them, worst first, each with a fix. The waste I hadn't noticed I was creating.
🔥 https://gyorgy.sh/blog/claude-code-expensive-habits

2. The tools I reach for
The agent, the terminal, the editor, the SSH client, the remote desktop, the local model, and the app that keeps the Mac awake through all of it.
💻 https://gyorgy.sh/blog/agentic-work-tools

3. Running Claude Code on local models
Two environment variables and the CLI runs against a model on your own machine.
🖥 https://gyorgy.sh/blog/claude-code-local-models

Last night a fleet of Claude agents rebuilt two of my sites while I slept.

One Telegram message to MyAgens deployed a Claude Code session on my Mac at home. I wrote the plan, an Opus 4.8 orchestrator held it together, Sonnet 5 sub-agents did the translations, Haiku handled the mechanical work. I approved the plans, went to sleep, and by morning there were around 500 server-rendered pages in 15 languages: keepresso.com in every language the app speaks, myagens.com in two.

For scale: getting my previous SaaS into 17 languages once cost us around 5000 EUR, spread over three years. This ran overnight, for the price of a lunch, inside a monthly subscription.

The translations are not perfect, maybe 97 percent there, and the next morning review was the real work. But the expensive part, wiring an entire codebase to be multilingual, is simply done.

Full devlog with screenshots:
➡️ https://gyorgy.sh/blog/claude-code-agent-fleet