Most MCP servers skip two of four tool safety hints, breaking confirmation heuristics
abhiramreddykalluru · reddit · 2026-10-26
MCP tools can carry four safety hints: readOnlyHint, destructiveHint, idempotentHint and openWorldHint. The author found most servers set at most the first two — but the last two tell AI apps whether running a tool twice is safe and whether it reaches outside services, which apps use to decide when to ask the user first.
Based on Anthropic's MCP building guide, the author added checks for these (plus clashing generic tool names like "search" and oversized tool lists) to an open-source tester. DeepWiki and GitMCP both skip the two labels; Context7 sets all four. Shown as tips, not scored.
More from coding & agent
- How to get more ambitious with evals as models approach superhuman intelligence — shyamalanadkat · 2026-10-26
- GenAICode: open-source toolkit puts one API over Claude Code, Codex, Copilot, Cursor and Gemini CLI — Codeblix_Ltd · 2026-10-26
- Claude rebuilds a Gaussian-splatted garden scene in Blender with Geometry Nodes — CSProfKGD · 2026-10-26
- Vocs: fully local macOS dictation with whisper.cpp streaming, short chunks decoded 8x faster — zenitsu--DS · 2026-10-26
- 50 agent runs, zero doc reads: five models pass tests but skip the documentation — ducha_aiki · 2026-10-26
- Every's 'frontier team' model: 100 AI experiments are useless if none land — every · 2026-10-26