Rohan Shield keeps API keys out of the LLM context window during MCP tool calls

In-memory privacy shield to prevent context leaks in MCP tools

Standard MCP tools run in cleartext at the host boundary, so credentials like sk_live keys and database passwords can leak straight into the model's context and inference logs. Rohan's open-source @rohan-protocol/mcp server evaluates tool arguments in isolated client RAM, strips zero-width Unicode prompt injections with a semantic firewall, and wipes witness buffers with deterministic zeroization. It ships as a public beta (v0.6.0) with a gasless Midnight relayer and a Zenodo/IEEE paper; native WASM ZK-SNARK proving is still under research.