A production-ready MCP trust scanner shipped in a week

A new open-source scanner called mcptrustchecker went live this week, and it does something most MCP users have been nervous about but not actively solving. It pulls the actual source code from npm or PyPI packages that claim to be Model Context Protocol servers, scans for tool poisoning and prompt injection patterns, then spits out a deterministic trust score from A to F. You can run it in CI pipelines to block dodgy packages before they touch production.

The timing matters. The same week this tool shipped, separate threads on X were raising red flags about MCP servers quietly loading thousands of tokens into context or acting as memory middlemen without clear disclosure. That is not coordinated, it is convergence. When multiple groups start building defences around the same risk at the same time, the risk is real and people are feeling it.

Why this lands now

MCP lets you plug third-party servers into your AI workflows to extend what models can do. That is powerful, but it also means you are handing those servers access to your prompts, your context, and sometimes your data. If a server is malicious or just poorly written, it can inject instructions, bloat your token spend, or leak information. Until now, most teams were either trusting package metadata or doing manual code reviews. Neither scales.

A deterministic scanner that reads source and gates deployment is the kind of guardrail that turns MCP from a proof of concept into something you can actually run at scale. If you are building an AI command centre that orchestrates multiple agents and third-party tools, this is the kind of supply chain check you want baked in from day one. The original repo is public, so you can fork it or run it yourself. It is early, but it is solving a problem that was not theoretical anymore.