Developer

How Alex Rezvov's OpenClaw Digest Bot Hit 16 Incidents in One Day

Built by Alex Rezvov · · Source: Blog

Quick answer

Alex Rezvov set up OpenClaw on a VPS to send a daily Telegram digest of Hacker News articles about AI agents. The post says it took 16 incidents, about ten hours and $1.50 in API calls to get the first working delivery.

Result: First working digest landed in Telegram about ten hours after install. That run took about 30 seconds and cost $0.03. Total API spend for the day was $1.50. The post does not confirm the setup kept running unattended.

What they built

Alex Rezvov set up OpenClaw to send a daily digest of Hacker News articles about AI agents to Telegram at 05:20 CET. The write-up on their blog is bylined Claude, a developer persona narrating in first person. According to the post, Alex mostly wanted to test OpenClaw configured the way it is meant to be, and the day produced 16 incidents before a digest arrived.

How it works

The post says OpenClaw runs on an existing Ubuntu VPS that also hosts Jitsi Meet, with the dashboard behind nginx, SSL and basic auth. Models come through OpenRouter: Mistral Small 3.2 for chat and DeepSeek Chat v3.1 for heartbeat and cron. The agent uses blogwatcher for RSS, exec for shell commands and web_fetch.

The schedule first lived in HEARTBEAT.md. The incidents the post describes include:

  • The schedule never fired. The heartbeat ran every 15 minutes and answered HEARTBEAT_OK each time, because Mistral Small could not reliably compare times.
  • A 271 message loop. After the file was changed to run the task unconditionally, one failed tool call had the bot send 271 Telegram messages. It ignored stop commands, and Alex ended it by killing the OpenRouter API key. The post puts the loop's cost at $0.42.
  • The thinking block trap. With thinking set to low, DeepSeek did its work inside the thinking block and output nothing or a fragment, while OpenClaw still reported delivered: true. One full 15 article digest was never delivered this way.
  • Bad output. Two of three generated Hacker News discussion links pointed to unrelated comments, the model kept asking for a file that did not exist, and control tokens leaked into Telegram.

The working version runs the digest as a cron job with thinking off. The post also recommends isolated sessions for cron jobs, timeouts of at least 180 seconds, anti-loop rules in SOUL.md, a budget cap at the provider, and checking Telegram yourself instead of trusting the delivered flag.

The result

According to the post, a cron run at 15:25 CET took about 30 seconds and cost $0.03. It scanned five RSS feeds, filtered 54 articles by date and delivered the top six to Telegram. That was about ten hours after install. Total API spend for the day was $1.50, and the post says it would have been $0.03 if everything had worked the first time. It does not confirm the next scheduled run. It ends with "We'll see if it fires."

How to build something similar

  • Use a real scheduler such as cron for timed jobs. Don't ask a small model to read a file and decide what time it is.
  • Set a spending cap with your model provider before the agent can run on its own, and know how to cut its API access fast.
  • Give the cheap model the chat and a stronger model the scheduled work, and turn thinking off for any job that has to produce a message.
  • Check the actual output in the channel after each test run. A success flag in the logs does not mean a message arrived.

The original post

Alex Rezvov · View on Blog

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases