Lessons from a LinkedIn MCP server that takes real actions
Calm_Ambassador9932 · reddit · 2026-11-13
The We-Link team shares design lessons from building an MCP server that acts on LinkedIn (search, connection requests, messages): confirm before irreversible actions; deliberately slow 10–60s human-paced execution with rate limits in the server, not the prompt; never put credentials in chat (separate secure sign-in link, removing the connector cuts access); explain destructive tools before asking for confirmation; and state limits in tool descriptions so the model says "I can't" instead of hallucinating. Open questions: where confirmation should live (server/client/model) and how to handle long-running actions in fast-response clients.
More from coding & agent
- Dev builds gear train demo in Rust with manually drivable gears more fun than the shader — thespite · 2026-11-13
- Plannotator passes 10k GitHub stars: one workspace to review and steer coding agents — backnotprop · 2026-11-13
- claude-trading-skills hits 3.8k stars: Claude Code skills for equity traders — tradermonty · 2026-11-13
- 10 open-source projects securing AI agent skills, as 26% of public skills show risky patterns — bibryam · 2026-11-13
- Cloudflare open-sources security-audit skill: multi-agent repo security audits, 71.5k stars — bibryam · 2026-11-13
- Chrome team's Modern Web Guidance hits #1 on GitHub 48 hours after 1.0 launch — rseroter · 2026-11-13