A company runs 17 Claude bots for $100 a month using a single Mac

Bypassing the API bill

A B2B distribution company in Korea has managed to run 17 Claude Code bots for a flat monthly fee of $100. While many AI automation projects involve high, unpredictable costs from per-token API usage, this setup uses a Claude Max subscription plan. This allows the company to perform various tasks without paying extra for every interaction.

The reason the API bill remains at zero is that the bots primarily run via the Claude Code CLI. Instead of wiring every bot to an API key, the owner only uses the API for three specific tasks. The other bots are built using plain Python to move data between databases, templates, and Slack cards. This setup is so simple that the owner noted the cost table is almost annoyingly short.

For the three bots that do require a model, they use a specific command line to request information. This method treats the AI as a simple input-output layer rather than a complex integration. As botsinproduction.com reported, the bot writes a prompt, receives a JSON response, and then moves on to the next task. This approach avoids the need for token accounting or complex key rotation ceremonies.

Technical guardrails and limits

To keep the bots reliable and safe, the owner uses specific flags in the command line. The --restricted and --strict-mcp-config flags ensure the bots cannot touch parts of the filesystem they aren't supposed to access. This prevents a chat summarizer, for example, from browsing unrelated files out of curiosity. Another important setting is --no-session-persistence, which forces every call to start with amnesia.

This amnesia ensures that the bot does not remember a previous customer when it moves on to a new one. The owner also enforces a strict JSON output format so that the system can easily parse the data without dealing with half-finished sentences. The owner also caps the number of tokens the model can return for specific tasks. For example, claim summaries are limited to 1,200 tokens because anything longer suggests the bot has gone off the rails.

There are trade-offs to this subscription-based model. The Max tier has usage limits that are 5x the Pro allowance on that tier. If the CLI login expires on the machine, all 17 bots can fall over at once like dominoes. The owner also learned that running database writes in parallel caused 300-odd database is locked errors and caused seven chats for the bot to simply give up on.

Hardware and maintenance

The entire infrastructure runs on a single, ordinary MacBook located in the office. There is no cloud hosting, no Docker containers, and no complex deployment pipelines. The owner uses a tool called launchd to keep a Python process alive, meaning the only requirement is that the laptop stays plugged in and does not go to sleep. When a bot goes quiet, the first thing the owner checks now is the Mac’s power log rather than the bot itself.

Maintaining this setup requires a human to handle the parts that are hard to automate. The owner spends about one hour a week on the bots now, though it took closer to ten hours a week during the first two months. This time is spent writing and rewriting rules, reading alerts sent to a direct message, and benchmarking different models to ensure accuracy remains high.

The owner started on the Pro plan for US$20 and moved to the Max 5x plan for US$100 when the number of bots multiplied. This tier has been sufficient for all 17 of them. The owner emphasizes that while the US$100 is a small cost, the human time spent writing rules and reading alerts is the biggest line on the bill by a mile. The goal is to have the MacBook do the work while the human simply checks the homework.