botler-agent: A Minimalist Personal Data Agent That Runs on an Old Mac

General-purpose Agent frameworks are heavy, expensive on tokens, and often too permissive for a long-running personal daemon. botler-agent shows that a five-tool, allowlist-only design can handle daily record-keeping and reminders at a fraction of the cost, on low-end hardware, without exposing the rest of the filesystem.

botler-agent strips the Agent down to a single tsx process that installs in seconds and runs comfortably on older hardware. It receives messages from chat platforms, routes each one to the correct data sub-project using a tiny routing prompt, then loads only that project's AGENTS.md to guide execution. The framework enforces a strict directory allowlist and gives the Agent exactly five tools — read, write, edit, run (existing scripts only), and schedule — with no arbitrary shell access. Every write is validated as legal JSON, and any change triggers an automatic git commit for backup.

Business logic lives entirely in per-project AGENTS.md files. Adding a new data domain means creating a directory and writing its rules; the framework source stays untouched. The author runs six projects covering vehicle maintenance, diet, daily logs, bookkeeping, English learning, and travel — all vertical personal data that chat-based input turns into structured, version-controlled records.

Token costs are kept low because each message spawns a fresh, stateless Agent. The routing phase uses only project names and summaries; the full AGENTS.md is concatenated only for the selected project. When routing is uncertain, the system asks the user to clarify rather than guessing. Scheduled reminders, a local WebUI, and health monitoring round out the optional modules, and all configuration lives external to the data directory so secrets never leak into version-controlled data.