Dev shares MCP lessons: human-only publishing enforced by omitting the tool entirely
mariobalca · reddit · 2026-11-20
The author built Tamable, a writing-and-publishing workspace whose MCP server exposes 27 tools to Claude, ChatGPT and Claude Code. Core rule: facts come only from the writer's own words, and only a human approves or publishes. Design choices that worked:
- No tool for the final step: agents can capture, draft, edit, submit for review and schedule, but approve/publish simply don't exist as tools — enforced by omission, not prompts;
- Rules live in server instructions and tool descriptions, so every client follows them, not just the one with a good system prompt;
- Admit uncertainty instead of guessing: unclassifiable ideas stay unfiled with an unfiledreason;
- Two entry paths: an OAuth connector for claude.ai/ChatGPT and personal keys for header-sending clients like Claude Code.
The author invites discussion on where to draw the agent-vs-human boundary.
More from coding & agent
- Your old CLAUDE.md is hurting new Claude models; author trimmed it to 36 rules-free lines — PrajwalTomar_ · 2026-11-20
- Open-sourced Claude AI team: PM, CTO, UX roles that challenge your product ideas — marilynika · 2026-11-20
- Open-source Intern harness: Opus 5.5 hands off to Hy4 Flash, 92.4% at 1/3 the cost — m4rkGS · 2026-11-20
- Gordon Wetzstein: coding agent in the loop boosts optical NN accuracy 35% to 96.8% — GordonWetzstein · 2026-11-20
- Optics as Code: Stanford agent tunes optical neural nets from 35% to 96.8% via hardware-in-loop experiments — GordonWetzstein · 2026-11-20
- Post-acquisition: Cognition's Poke agent gets its own computer and a chat interface — marvinvonhagen · 2026-11-20