Free toolRuns in your browser · nothing is sent or stored
MCP security checker
Twelve questions about an MCP server or AI agent you run or plan to run. Get a risk rating and a prioritised list of what to fix first.
Your MCP server or agent
Pick the answer closest to how it is set up today, not how it will be. Unsure is a valid answer.
The questions follow the security guidance in the Model Context Protocol specification and what I learned running my own MCP server. Building a server? Start from the MCP server generator, or have me build and review it.
How it works
01
Answer twelve questions
They cover the account the server uses, what it can change, confirmation for irreversible actions, how it is reached, prompt injection, what it returns, secrets, logging, limits, where the code comes from and testing.
02
Get a rating
Any answer that could do serious damage on its own makes the rating High. Several medium gaps make it Medium. The meter shows how the gaps add up.
03
Fix from the top
Each issue comes with a concrete fix, ordered from most to least serious. Re-run the check whenever you add a tool or connect a new system.
Questions
Does a Low rating mean my MCP server is secure?
No. This is a self-check, not an audit or a certification. It only knows what you told it, and it cannot see your code, your network or your accounts. A Low rating means your answers describe a careful setup.
What is the biggest risk with MCP servers?
Usually the account the server runs as. A server can do whatever its credentials allow, whatever its tool descriptions say. After that come irreversible actions without a person in front of them, and tools that read text written by outsiders in the same session as tools that can write.
Do the read-only and destructive annotations protect me?
No. Tool annotations are hints for the AI client, and the MCP specification says clients must treat them as untrusted unless the server is trusted. What protects you is code that cannot write, and an account that is not allowed to.
What is prompt injection in an MCP context?
When a tool returns text someone else wrote, such as an email, a ticket or a web page, that text can contain instructions aimed at the AI. If the same session can also send or delete things, those instructions can turn into actions. Keep reading outside text and writing apart, or put a person in front of writes.
Is anything I answer sent to SlickLab?
No. The check runs in your browser and nothing is sent or stored.
The next step
Want a person to look at your result?
A free tool gives you a first answer. The MCP Server Development service turns it into a plan for your business, with a fixed quote before any build.