How Refract built a remote MCP server with write access to social accounts
duma2404 · reddit · 2026-11-16
The team behind Refract (a social media scheduler) shared how they designed auth and permissions for their hosted MCP server (mcp.refract.social, listed in the official MCP Registry), a useful reference for anyone building agents that take real actions.
- Auth: OAuth discovery + dynamic client registration + PKCE; works in Claude, ChatGPT, Cursor and VS Code with no token pasting.
- Tiered permissions: users pick "draft/send for approval" or "schedule/publish" at connect time, mapped to existing team roles so the agent follows the same approval rules as a person.
- 19 tools: read tools for context, find-slots for posting times, and check-post, which returns per-network issues with block/warn/info severity so the model can self-fix before requesting approval.
- Network quirks live in tool descriptions: e.g. TikTok's privacy level must come from the user — "never choose it for them" — which worked better than any client-side prompt.
- The author also asks how others gate destructive tools like publish or delete.
More from coding & agent
- Dev switches to cheap AI model during 5-day maintenance, ends up with buggy codebase and credential threats — emeka_boris · 2026-11-16
- Localix Terminal: open-source Rust terminal with project workspaces and agent integration — Awkward-Let-4628 · 2026-11-16
- Websites increasingly block AI agents, and the agentic web may fragment further — signulll · 2026-11-16
- Firms saving $100k with AI-written code may pay $500k to fix collapsed architecture — PAstynome · 2026-11-16
- Redditor builds a Doom-engine Linux desktop environment where directories are map rooms — triotune · 2026-11-16
- ai-rulez v5 serves repo skills and scoped context to coding agents on demand over MCP — Goldziher · 2026-11-16