MCP Snitch

MCP Snitch doesn't rely on static rules alone. It runs an LLM check over the input and output of tool calls to spot suspicious patterns as they happen. That means it can catch odd behavior a plain allowlist would miss. Think of a normal-looking request that carries something unusual in the parameters. So which calls does it trust? Only the ones your policy clears.

Sensitive keys get moved out of plain config files and into a secure keystore automatically. When a tool tries to read environment variables holding those keys, the call is blocked and the key stays out of reach. If you've ever worried about an MCP server quietly scraping your credentials, this is the feature aimed at that.

You control which MCP tools are allowed to run, down to the individual call. New or unfamiliar calls can require explicit approval while known-good ones keep working, which is the difference between a blanket ban that kills your workflow and a policy you can actually live with day to day.

Whenever a new tool call shows up, you get a notification with three options: always allow, allow once, or block. Native notifications mean you don't have to sit in the dashboard waiting. Decide once. The app remembers your choice.

MCP Snitch keeps a whitelist of verified MCP servers, so confirmed ones run smoothly while anything new is treated with suspicion. Over time this builds a clear picture of which servers your environment actually trusts. It's a small thing. It adds up.

Every tool interaction lands in a logged audit trail, with the input, output, and the decision that followed. When a call gets blocked or flagged, you can see why. That's how you audit MCP tool calls without digging through logs by hand. Reviewing that history is how you tune the policies without guessing.