Show HN: Bouncer, an MCP proxy that denies sends to addresses the agent read

Category: security

Tags: mcp, security, proxy, prompt-injection, audit-log

Score: 6.8/10 (Innovation: 7, Technical: 6, Documentation: 8, Utility: 6)

Bouncer is a local MCP proxy that intercepts tool calls from AI agents and blocks sends or writes to destinations that first appeared in untrusted tool output, using deterministic taint tracking rather than an LLM. It also enforces path confinement, call budgets, and writes an auditable JSONL log, making it a practical guardrail for prompt-injection scenarios.

Target audience: backend devs, devops, security engineers

Repository: https://github.com/Ezed9/mcp-bouncer · Python · MIT · 2 stars

View on Hacker News