Know what an MCP server does before your agent trusts it.
An MCP server is a plug that gives an AI agent new tools. Aevrin runs the server in a safe sandbox, reads every tool it offers, and grades it from A to F.
Scan notes-sync before I add it
Aevrin- Started it in a sandbox· a safe, throwaway copy
- Read every tool· 6 tools
- Found hidden instructions· in sync_notes
Decision: Block. Your agent will not use it.
The problem MCP Scanner solves.
An MCP server is code your agent runs for you. It describes itself in its own words, and those words can lie. Adding one on trust is how hidden instructions, copycat tools and bad packages get in.
- 1
Scan the server
- 2
Grade the risk
- 3
Allow, ask or block
What goes in
- The MCP server and its settings
- Each tool's name and description
- What each tool accepts and may do
- Its source code
- Its package details
- Its libraries and their versions
What comes back
- Each problem it found
- A grade from A to F
- A risk score from 0 to 100
- Risky permissions and attack patterns
- Package and secret-handling risks
- Allow, Require approval or Block
Why you can trust what it says.
A clear decision
A and B mean allow. C and D mean ask a person first. F means block. Your agent can follow that rule directly.
Honest about what it missed
If a server hides its tools, the scan says it could not check them. An unfinished scan is marked incomplete and gets no grade. It is never called clean.
Uses a shared checklist
Each problem is matched to the OWASP MCP Top 10, a public list of the biggest MCP risks. It reads the same in the dashboard, the terminal and CI.
The other five checks.
Put MCP Scanner to work.
Start on the free plan. No card, and nothing renews on its own. Or add Aevrin MCP and use it from your agent.