Lessons from building 5 deterministic MCP servers for finance/data tasks
Ltx_23 · reddit · 2026-11-15
After months building MCP servers for boring-but-critical tasks (bank statements, accounting exports, CSV cleaning), the author shares three lessons:
- Keep tools narrow: one tool doing one exact job beats a "do everything" tool — models pick them far more reliably.
- Return structured results plus a short summary so the model doesn't re-interpret the numbers.
- Remote + OAuth makes onboarding 10x easier than asking users to edit JSON config.
The code is source-available (BSL) on GitHub (ArceeDev) and on PyPI.
More from coding & agent
- Garry Tan: Only ~1% of engineers let agents ship features — it should grow 100x — garrytan · 2026-11-15
- Prompting regime is back: scaffolding and context now separate vastly different results from same models — Afinetheorem · 2026-11-15
- MarkItDown becomes Microsoft's most-starred GitHub repo with 196k stars, overtaking VS Code — SaleemaAmershi · 2026-11-15
- LMMs Lab open-sources The Harness: one unified framework for computer, robot, and game agents — liuziwei7 · 2026-11-15
- shadcn authors designed UI libraries by inspecting model weights, making agents better — cramforce · 2026-11-15
- Google launches Gemini Decisions API with Gemini Decision 1, zero output-token billing — sebkrier · 2026-11-15