While most enterprise AI case studies drown in cloud infrastructure costs and API token burn rates, a new report from a Korean B2B distribution company offers a starkly different reality. OI, the company’s operations lead and self-described non-developer, runs 17 distinct Claude Code bots on a single MacBook for exactly $100 a month. The entire operational stack relies on a Claude Max 5x subscription, with zero spend on the Anthropic API. This isn’t a theoretical demo; these bots handle customer complaint filing, monthly statement emails, Slack error monitoring, and dynamic pricing adjustments daily.

The CLI as an Integration Layer

The cost efficiency stems from a counter-intuitive architectural choice: avoiding the API entirely for most tasks. Only three of the seventeen bots actually interact with a large language model; the remaining fourteen are plain Python scripts handling database inputs and template outputs. For the three AI-dependent bots, OI bypasses the SDK and token accounting by shelling out directly to the Claude Code CLI. The command claude -p --model --restricted --no-session-persistence --strict-mcp-config --output-format json serves as the entire AI integration layer. This approach leverages the existing subscription’s flat rate, turning the CLI into a cheap, stateless inference endpoint that accepts prompts via stdin and returns JSON via stdout.

Guardrails Over Generosity

OI’s setup prioritizes strict guardrails over model generosity, a necessary discipline when running unattended automation. The --restricted and --strict-mcp-config flags prevent bots from accessing filesystem tools they don’t need, while --no-session-persistence ensures every call starts with amnesia, preventing data leakage between customer interactions. For the chat summarizer, an API fallback exists but remains disabled because the CLI handles the load within subscription limits. Crucially, output is capped at 1,200 tokens and limited to eight items per response. If a summary exceeds these bounds, it’s discarded as a transcription error rather than a summary, forcing the system to retry rather than hallucinate. This strict validation prevents the "long heartfelt conversation" problem that often plagues automated agents.

The Hidden Cost of Human Oversight

The $100 hardware and subscription cost masks the true operational expense: human time. OI spends roughly one hour per week maintaining the system, a figure that dropped from ten hours during the initial two-month setup phase. This time is consumed by refining logic rules—such as distinguishing between missing items and quality complaints—and monitoring DM alerts for bot failures. The infrastructure is equally minimal, running on launchd on a MacBook that must never sleep. OI notes that laptop sleep cycles caused more initial failures than any prompt engineering issue, requiring a shift to full paths in configurations to avoid PATH resolution errors common in service environments.

Key Takeaways

  • Subscription Arbitrage: Using the Claude Code CLI via subscription can eliminate API costs for low-volume, high-frequency tasks, provided usage stays within tier limits.
  • Statelessness is Safety: Enforcing --no-session-persistence prevents context contamination between unrelated customer interactions, a critical feature for B2B operations.
  • Hardware Simplicity: A single laptop with launchd is sufficient for 17 production bots, eliminating the need for Kubernetes, Docker, or cloud instances.
  • Output Caps are Correctness Tools: Limiting token output prevents models from drifting into transcription mode, ensuring summaries remain concise and actionable.

The Bottom Line

This case proves that enterprise-grade automation doesn’t require enterprise-grade infrastructure; it requires rigorous constraints and a willingness to treat the model as a dumb pipe rather than a conversational partner.