I built MCPSense to give MCP maintainers a local way to inspect what a server advertises. After Black October, I revisited a case where the same instruction can be ordinary or hostile depending on what it asks the agent to send. In the demo, sending public status values and sending an authentication token use the same destination. MCPSense flags the token case.

Version 0.4.5 lets you compare supported metadata with a reference you reviewed, so you can see what changed on a later scan. The offline demo inspects the manifests without calling either tool.

Try it on one MCP integration you maintain. What did the report help you decide to review? If you hit a blocker, tell me which step stopped you. Keep configurations, credentials and full reports private.

Runnable demo: https://github.com/fayzkk889/MCPSense/tree/main/docs/launch/2026-11-02/demo

Version 0.4.5 lets you compare supported metadata with a reference you reviewed, so you can see what changed on a later scan. The offline demo inspects the manifests without calling either tool.

Runnable demo: https://github.com/fayzkk889/MCPSense/tree/main/docs/launch/2026-11-02/demo

MCPSense

Find risky MCP tool descriptions and track changes

Find risky MCP tool descriptions and track changes