Ultra Lab 的 AI agent 用什麼模型？

先用本機 Ollama 上的 qwen 模型，本機不通才改走 OpenRouter。排程由 systemd 定時任務觸發，出錯推 Telegram 告警。

How much do the agents cost to run?

Model spend is on the order of a few US dollars a month. The local model handles most requests; OpenRouter is the fallback.

Ultra Lab 的 AI Agent：這頁只列現在真的在跑的

Q: Which AI agents does Ultra Lab run today?

Three parts are visible from outside: Xiao Q (@Ultra_Agentbot), an internal Telegram assistant on a local qwen model that sends a daily product report and posts one note a day on Moltbook; the Discord community bots, which post daily AI news, run the onboarding quest, level members by activity and announce new blog posts; and a watchtower that alerts on Telegram when a scheduled job misses its run. Day-to-day operations are run by an AI operations agent in a COO role and headless workers per product line; people make the decisions.

公司日常營運由一個 AI 營運代理（營運長角色）和各產品線的無頭工人執行，決策由人做。下面是你從外面看得到的幾個。

名冊

只列實際查過、確認還在執行的元件。停掉的就從這裡拿掉。

小Q（@Ultra_Agentbot），Telegram 內部助理：跑在我們自己電腦上的本機 qwen 模型，只替老闆一個人工作，不對外接受對話。每天早上推產品日報，每天在 Moltbook 發一篇隨筆。
Discord 社群機器人：每天在 #general 貼一則 AI 新聞、新成員完成入門任務後給身分組、依每日活躍度從 Lv1 升到 Lv4、部落格有新文章時發通知。
監控哨：檢查定時任務有沒有按時執行、每支看門狗的心跳（包含自己），每天早上 8 點推 24 小時健康摘要，出事直接推 Telegram。

技術架構

排程由 WSL2 Ubuntu 上的 systemd 定時任務觸發。模型先用本機 Ollama 上的 qwen，本機不通才改走 OpenRouter。模型花費每月數美元等級。

頁面上的像素辦公室是名冊成員的示意動畫，不是即時畫面。

What runs today (English)

Xiao Q (@Ultra_Agentbot) is an internal Telegram assistant on a local qwen model; it sends a daily product report and posts one note a day on Moltbook. The Discord community bots post daily AI news, run the onboarding quest, level members by activity and announce new blog posts. A watchtower checks that scheduled jobs actually ran and alerts on Telegram. Day-to-day operations are run by an AI operations agent in a COO role and headless workers per product line; people make the decisions. Models: local Ollama first, OpenRouter as fallback; model spend is a few US dollars a month.