Where should approval for MCP write tools be enforced? A Reddit MCP test
Special_power_90 · reddit · 2026-09-24
A developer tested a local Reddit MCP server that exposes read-only tools (browse, search, read posts) alongside side-effecting write tools (reply, post, vote). Writes only felt safe after someone reviewed the exact text and explicitly approved that single action.
For anyone building agents with side-effecting tools, the core question: should the approval boundary be enforced client-side, server-side, or both? And what should be logged so the executed action can be audited against what was approved?
A practical discussion of MCP permission governance: when read and write tools share one server, per-action approval granularity and execution auditing are gaps the ecosystem hasn't solved.
More from coding & agent
- A GitHub collection of project-based C tutorials, from computer architecture to OS internals — tetsuoai · 2026-09-24
- Codex Computer Use driving WeChat via iPhone mirroring dodges account bans — vista8 · 2026-09-24
- GPT-6 Astra Fixes a UI That Even Opus 5.5 Couldn't Crack — haltakov · 2026-09-24
- Kimi client patches arbitrary-command vulnerability via local embedded server — KimiDevs · 2026-09-24
- Codex computer use drives iPhone via Mac mirroring, fast and undetected — op7418 · 2026-09-24
- Kimi Code Desktop 1.0.3 adds built-in browser for local HTML files — KimiDevs · 2026-09-24