What 17 Claude Code Bots Cost Per Month: US$100 (Real Numbers)
B · Useful, limited First-hand practitioner post by a non-developer operations lead at a Korean B2B distributor with a full cost table and named failure modes; useful but self-reported by one pseudonymous author.
Summary based on the full text·Original: about 8 minutes to read
In short
A blog post (21 October 2026) by "OI", who runs operations at a B2B distribution company in Korea and is not a software engineer. Since 2025 seventeen bots built with Claude Code log complaints from chat transcripts, send monthly statements, watch Slack for errors and price products. They run on one office MacBook and call the Claude Code CLI under a US$100 subscription instead of the API. The largest real cost is the author's own time.
Key numbers
- US$100monthly cost of the Claude Max 5x plan that runs all 17 bots; API spend US$0
- 3 of 17bots that call a model at all; the rest are plain Python
- about one hour a weekthe author's time on the bots now, against about ten in the first two months
- more than 40hand-written rules in the claims logger, each from a wrong row
As reported by the source.
English·Web page·opens botsinproduction.com
What it reports
- Bill: Claude Max 5x US$100; API US$0; hardware, Slack, Google Workspace and Grafana add nothing extra.
- Only three bots talk to a model (chat summariser, photo matcher, error watcher); the other fourteen are database-in, template-out scripts.
- Model calls shell out to the CLI with restricted tools, no session persistence and JSON output, so failures are explicit.
- Guardrails: output capped at 1,200 tokens and at most eight items per summary; summaries run three at a time, sheet writes one at a time after 300-odd "database is locked" errors.
- Hidden cost: about ten hours a week in the first two months, about one hour now, spent on rules, alerts and re-benchmarking after model swaps.
- Failures: an expired CLI login took all 17 bots down at once; a sleeping laptop looked like a bot fault; launchd not loading the shell PATH broke bots silently.
- Every bot action passes a Slack approval button before it ships.
How they got it
The author's own logs, git history and read-only queries of live systems, as stated in the post. Drafted with Claude and checked by the author. No independent verification.
Limits
- One company, one pseudonymous author; no accuracy or business outcome figures.
- The post is inconsistent on which bots use a model: it says three do, then lists the claims logger, statement sender and price bot among CLI users.
- Running company automation on a personal subscription may conflict with plan terms; the post does not address this.
- No security review of a single unattended laptop holding company data.
Why it matters for companies
Most of the value here comes from plain scripts with a model at a few narrow points, plus a person writing rules from real errors. The single shared login and single machine are concentration risks that IT would not accept at scale, so bots like these need an owner, service credentials and a migration path once they matter. The honest hours line is the useful number for anyone sizing citizen-built automation.
Summary based on the full text. Read the original for the full argument and data.Read the original on botsinproduction.com