Self-hosting Hermes Agent for $3.50 a month: Zeabur, Telegram and DeepSeek Flash

Summary

We self-hosted Nous Research’s Hermes Agent as an internal news-reading agent for a 30-person team — deployed through Zeabur onto a $3/month Tencent Cloud server, plus roughly $0.50/month in DeepSeek Flash usage, delivered straight into Telegram. About $3.50 a month, total.

Problem

The team wanted something that reads local business, marketing, ad-platform and AI news every day so nobody has to, and pushes anything relevant into Telegram. A general-purpose AI subscription at $20/month is built for open-ended chat, not one narrow scheduled reading job — paying for a whole subscription just to run this would have been the wrong tool for the size of the job.

Architecture

Hermes Agent runs on a Tencent Cloud VPS, deployed through Zeabur. It reads an RSS digest of local business, marketing, ad-platform and AI news, then hands scoring and drafting off to OpenRouter, split between DeepSeek’s Pro and Flash models by task — cheap Flash calls for routine scoring, Pro kept for anything that needs more judgement. Telegram is the delivery gateway, so the digest lands where the team already reads everything else.

The Hard Parts

This is the one part still missing a real answer: what actually broke while wiring up the Zeabur deployment, the Telegram bot, or the OpenRouter/DeepSeek routing. It’s left blank here rather than guessed — Rafli’s first-hand account goes in before this ships.

Results

The agent runs for about $3.50 a month — a $3/month Tencent Cloud server plus roughly $0.50/month in DeepSeek Flash usage — against what a general-purpose $20/month AI subscription would cost to run this one job alone.

Stack

Tencent Cloud hosts the VPS, Zeabur handles the deployment, OpenRouter routes between DeepSeek’s Pro and Flash models by task, and Telegram delivers the result.