MCP tool descriptions silently change — and none of the 4 major AI coding tools warn you
abhiramreddykalluru · reddit · 2026-11-08
The author audited docs of Claude Code, Cursor, VS Code and Windsurf: none say whether users are re-prompted when an already-approved MCP tool's description changes. Yet "always allow" approves one version, and descriptions are effectively prompts.
Key findings:
- All four tools offer per-tool toggles, run-time confirmation, and "always allow"
- Claude Code's docs only mention auto-refreshing the tool list on listchanged, with no user notification
- Monitoring public MCP servers hourly for over a week, descriptions gained sentences far more often than lost them; one server's tools changed 15 times in 8 days
The author (building an open-source MCP tester) asks whether a simple alert suffices or tools should be blocked until re-approval.
More from coding & agent
- Is Hardware the New Software? Redditors Weigh Building Physical Products with ChatGPT — themegapint · 2026-11-08
- Opus 5.5 reportedly runs a native idea-level actor-critic self-play loop — sytelus · 2026-11-08
- Harbor turns one: 200T tokens/week, 10M+ Terminal-Bench runs, 1B+ rollouts — BenBlaiszik · 2026-11-08
- GrokBot + Notion pipeline brings job data to near real-time, revealing AI salaries of 300k-1M RMB — huangyun_122 · 2026-11-08
- OpenClaw explores "Slide to Auto-Merge" — new UI/UX patterns for the agent-first era — vincent_koc · 2026-11-08
- New AACL 2026 survey maps the end-to-end training pipeline of LLM-based agents — kalyan_kpl · 2026-11-08