Dev open-sources MCP proxy that gates every tools/call — and documents its three security limits
ceyptoslim · reddit · 2026-11-22
A Reddit developer open-sourced (AGPL-3.0) an MCP proxy that sits between an agent and upstream servers: every tools/call is decided before forwarding, and a denial returns an error carrying the rule and audit-record hash without ever contacting the upstream.
Three documented boundaries the author wants critique on:
- Sole-path assumption: mediation only works when the proxy is the agent's only route (exit class EX-4); closing it requires egress configuration outside the repo.
- tools/list passes unfiltered, so the model sees tools it will then be denied; filtering by tenant policy risks changing the model's plan, not just permissions.
- It records decisions, not outcomes — an allowed-but-failed call looks identical to an allowed-and-successful one.
Biggest limit: decisions are at tool granularity, so an injected call using an allowlisted tool with different arguments passes the gate. The author pre-registered this expectation and asks the community about threat-model gaps and how HTTP/SSE transports affect the sole-path assumption.
More from coding & agent
- Browser dragon-in-your-hand demo built with Three.js, MediaPipe hand tracking and WebRTC — techartist_ · 2026-11-22
- Developer Praises Google Antigravity's Agent Swarm UX — majidmanzarpour · 2026-11-22
- Store Scraper MCP exposes App Store and Google Play data to AI agents — modelcontextprotocol · 2026-11-22
- etapa MCP offers a free AI cycling coach for beginners — modelcontextprotocol · 2026-11-22
- Claude Code plugin turns hooks into persistent memory shared across Codex and Cursor — blaizedsouza · 2026-11-22
- Brandon Galang: As Models Get Capable, UX Scaffolding Melts Away in Favor of Context Flow — brandongalang · 2026-11-22