L Plo 🧲 (@plubtcm) header
L Plo 🧲 (@plubtcm) profile picture

I automate my work with AI agents and open source the tools. Local models, homelab, automation. AI builder nights in Amsterdam. https://nitter.gq/t.co/V2z9QKb9EL

Joined December 2017
Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)Photo by L Plo 🧲 (@plubtcm)
I still write API keys to a local env file, and I keep them out of the chat. If a tool needs the key in the prompt, I will not use that tool. Where do your agents get their secrets?
19
"Just install this agent, you're done." I counted what my agents actually knew. 27 rule files. 34 duplicated skills. 7 memory silos that could not see each other. The hard part was giving every CLI the same skills and memory, with a review gate.
2
93
Do you have a local model on disk that still runs if every API is down? If yes, reply with the model and how much RAM it needs. If no, what is stopping you?
4
2
42
"Just add another MCP server, you'll be fine." I keep the number of MCP servers on one agent small. Each new server is one more thing the agent trusts. What's the limit you actually use?
1
17
I host AI builder nights in Amsterdam every two weeks. If you build with agents or local models and you're in the city, come say hello. Reply if you want the next date.
1
36
1/ I needed one isolated agent runtime per customer, each in its own container. 6 steps, from a Telegram bot to a paired container. The last post is the limit list and the repo.
1
23
5/ Steps 5 and 6: they DM their new bot, get an 8-character pairing code, and paste it back into the storefront. The storefront auto-approves. The bot then replies normally.
1
1
6/ What this can't do yet: no billing, no encryption-at-rest for secrets in the control-plane database, and no idle reaper. A shared box is still the common setup. The flow is here: github.com/lpsmurf/hfsp-agen…
1
I built an 8-bit rental-store UI that browses Claude Code skills and top GitHub repos. It runs in the browser. You flip through the shelf like you would at a video store. What would you put on that shelf first?
2
39
When an agent wants to change a skill or a memory file, who approves it? On my setup, agents may write. A human promotes. Changes ship as a PR I review. Reply with how you handle that.
1
1
51
What decides which model does the work in your setup? I have a cost router in shadow mode. It logs what could have moved, and changes nothing until I trust it. Do you pick the model yourself?
19
The usual stack would have given me three alarms. This folded them into one: Demo gateway is down, taking 2 more with it. The two were Demo app and Demo API. The other 14 targets stayed up. Every target names a parent. github.com/lpsmurf/hermes-ho…
21