How do you manage MCP servers across multiple agents?

Here piped.video/watch?v=6eBSHbLK…

Time to sleep, but new reset in 25 mins  🥳

What hobbies do you have? To provide my brain enough space to grow and not be tied only to work, I started growing a tangerine. Next is lemon and orange, which will be bit challenging in our climate Tangerine going strong, still bit cold days with couple morning frosty so keeping it inside by the window.

America wasn't discovered, it was built. Did you watch The men who built America? I highly recommend it. I see AI comparing to steel from that time and it is changing how we live and work. Some will earn big buck with it, some will fail, but in the end our lives will be improved. Oil, steel, electricity, internal combustion engine, internet, mobile phone, AI. Probably missed some, for example I don't see crypto currencies on this level as human changing tech(prove me wrong?) Curious to hear your opinion on this @ivanburazin

Do you track X analytics? Why yes? Why not?

morning standup with your gang

Patching actually feels good.

Next challenge, access the dashboard and continue running agent sessions from any device

Making solid progress. Added "lui dashboard" command that opens a dashboard with all details about VMs, running processes, and agents. The CLI is so far agent agnostic, you can use Claude Code, @OpenAICodexCli, the Codex app with SSH connection, or any other agent.

Easy way to spin durable VMs locally for isolating projects, and running agents safely in full isolation? Working on a CLI for that. I call it Lui, and currently it mounts the selected folder to VM, exposes routes through Caddy so you can have multiple full-stack projects locally (isolated dbs and containers on the same ports while keeping codebase locally, for example, @supabase or @makerkit_dev). Went a step further and patched the Codex App(as I really like working with it) to have VMs directly in the UI. Still work in progress.

VM is running all the time, you can SSH into it, stop, start, etc, it has a state. This approach takes more resources. Ephemeral agent isolation exists only during the agent session(CC, Codex etc), after that it is cleaned. Lighter on resources. You can read this on how Claude Cowork isolation works aaddrick.com/blog/claude-des…

Couple things inside CLI: - it bootstraps a Lima VM and mounts project folder, forwards necessary ports, installs Codex stuff - Codex App is patched so it can see if folder is mounted and where - Caddy is making pretty localhost urls - You use Codex as normal but sessions are isolated inside VM This way you can run multiple projects fully isolated.

You could make it work with VMs too. For local dev I built a CLI that spins up Lima VM per project. Then worktrees are isolated inside it and you can use Claude Code or Codex CLI directly in VM. Currently making Codex app work this way too.

Run into some k8s issues with resource limits. Need to say that @k8slens is an awesome piece of software, as it helped greatly. Also, as I recently went deep dive into the awesome world of k8s, this is a free youtube video I would recommend everyone to watch piped.video/watch?v=MTHGoGUF…

Not everything is down :)

Hey everyone, would like to introduce you to sherpa.sh! A few months ago, I joined a friend in a pretty ambitious project. We are building a non-serverless Next.js (and a couple more frameworks, too) hosting platform, including scalable compute and high-performance CDN. Our goal is to be a more approachable, friendlier alternative to big and expensive hosting platforms. We’re small, fully bootstrapped, more personal, and handle all the server infrastructure, while giving you much more resources for the same price. You can read more about our architecture here docs.sherpa.sh/infrastructur…. On the app part, we are using NextJS 15 with Supabase in a Turborepo setup. We have a free plan if you want to try it out (still early days, so expect a few bugs). 😅 I will be sharing our interesting journey here, coupled with some of my smaller vibe-coded projects.

After fiddling with the cartoonize photo app for some time, I am niching down into the pet niche. Complete UI is again done in Lovable, now wiring up the logic and moving the app to NextJS. Check the UI here preview--toonilo.lovable.app… and remix it eventually. I am still impressed by how good @lovable is at crafting the UI.

A new policy was added into ChatGPT to avoid transforming faces or people. But a workaround is to use gpt 4o vision to describe the image, then feed 4o image gen with image and that description, setting description to be the primary source of information. Here is the exact gpt 4o vision prompt I use to describe the image.

Initial prototype was done using @lovable, which is a fantastic tool for building MVPs. After initial work on UI, I moved to Cursor. The first task was to refactor to NextJS app and continue from there. For database I use Supabase. For file storage Supabase Storage. There is no auth or payments at this point; it is free to use, so no auth. Only auth-related thing is a token when image is generated. For hosting I used sherpa.sh, which is a new platform for hosting NextJS apps.

So I was cooking something, lol. I don't remember the last time I shipped something. When we got 4o image generation, it looked incredible to me. As we don't have an API yet, it bugged me how to handle it. I saw creative ways, so I made my own "API" using a custom browser extension. Yes, it is a chatgpt wrapper, yes, there are many of them. I don't care. I just wanted to write down that I shipped something. Why am I shipping it now? Because of @thepatwalls email I received yesterday 😃 Give it a go: cartoonize.photo/. It is probably buggy, and there will be issues. Let me know how it goes and if you have any comments. I tried to avoid any copyright issue by not mentioning any brand names, etc.

Just converted a small MVP app made with @lovable into a NextJS app using @cursor_ai agent mode with Sonnet 3.7 Max. It went smoothly and took around 10 minutes, with some correction back and forth. It could probably be a one-shot with correct prompting, I used the most basic prompt. I would say this is awesome, Lovable for crafting awesome UI, NextJS if you need quick prototype with backend included. And all ruled with Cursor.

So, after years of thinking to build something on my own I finally decided to start. I am a long time wantrepreneur, started a lot of side projects, but then I usually stop, not sure why. Now I want to try and be more accountable and connect with other developers and marketing people on X. Have multiple ideas but will start with my own problem. For a physical business I co-own, we do a lot of carousels on Instagram. I am building a carousel editor that will allow for non-design people to develop and manage beautiful carousels. Did some research, and there are multiple solutions, but I decided to go with my own for the sake of learning and publishing something online. I will post my progress here. For tech stack I am using React, will pair it with NextJS for routing and simple backend, Zustand for local state(editor), Supabase for database etc.