Bouncer: A Deterministic Local MCP Proxy to Prevent Prompt Injection
eccentric_ez · reddit · 2026-08-12
Developer Ezed9 open-sourced Bouncer on GitHub, a local proxy designed to secure MCP servers. It gates tool calls using deterministic contracts rather than relying on an LLM in the request path.
Key Features:
- Schema Pinning & Budgets: Wraps MCP servers 1:1, enforcing argument constraints and call budgets.
- Taint Tracking: For high-risk actions like sendemail or sharefile, it checks the destination against provenance. If the recipient originated from untrusted tool output (e.g., a poisoned page), the request is denied. Trusted destinations are allowed, while unknown ones prompt for human approval.
The tool was benchmarked against AgentDojo's workspace suite, successfully catching bypasses in the developer's own code. Pre-configured packs for common servers like Gmail, Slack, and GitHub are available in the repository.
More from coding & agent
- Grok 4.6 available in Cursor and API with 2x usage for the first week — Daniel_Farinax · 2026-08-12
- Ref Raises $4M to Fix 'Velocity Sickness' in AI Coding Agents — round · 2026-08-12
- Practical Tip: Run LLM Agent Experiments Free via Kaggle CLI with T4 GPUs — mariofilhoml · 2026-08-12
- Dev Changes 'Proceed' Button to Encouraging Text to Please AI Agents — doodlestein · 2026-08-12
- DeepLearning.AI and JetBrains Launch Free Short Course on AI Coding Workflows — DeepLearningAI · 2026-08-12
- mini-swe-agent is the Best Harness for NVIDIA Nemotron on Terminalbench — OfirPress · 2026-08-12