Find what's actually
exploitable in your
MCP servers.

krysta-probe scans every MCP server on your machine for command injection, path traversal, credential exposure, and tool poisoning — then confirms real findings with live, safe exploit testing, so you're not chasing false positives.

$ npx krysta probe --deepCopy
Open dashboard See how it works
11
Vulnerability patterns
3
Detection layers
npx krysta probe --deep
Scanning for MCP servers...
 
Found 2 MCP server(s)
 
filesystem (Stdio)
Missing authentication
Potential path traversal risk
 
shell-server (Stdio)
Confirmed command injection
Environment variable dump
 
Risk score: HIGH
Findings: 4 (1 confirmed exploit)
Report: ~/.krysta/krysta-report.json

Real output structure from an actual scan. Server names shown here are illustrative.

How it works

Discover

Find configured servers

krysta-probe reads your Claude Desktop, Cursor, and VS Code configuration to find every MCP server you have set up — no manual list required.

Test

Check and confirm

Each server is checked against known vulnerability patterns at three layers — configuration, source code, and live protocol calls — with safe test payloads confirming real exploitability where possible.

Report

Get an actionable report

Findings are ranked by severity and confidence, with specific remediation steps and an auto-generated policy file you can apply immediately.

What it checks for

Ten documented vulnerability categories, drawn from known MCP-specific attack patterns.

Command injection

Tool schemas that allow arbitrary shell execution through unvalidated string parameters.

Path traversal

File read/write tools that lack path validation, allowing access outside the intended directory.

SSRF

URL-fetching tools with no allowlist, allowing requests to internal services or cloud metadata endpoints.

Credential exposure

API keys or tokens hardcoded in server environment variables or configuration.

Hardening bypass

Command allowlists that validate only part of an input string while executing the whole string.

Tool poisoning

Hidden instructions embedded in tool descriptions intended to manipulate the calling agent.

Supply chain risk

Servers pinned to floating versions, pulling unreviewed code on next install.

Unauthenticated endpoints

Network-exposed servers with no token or session verification.

Questions

Scan your MCP servers

One command, runs locally, no signup required to try it.

npx krysta probe --deep
Create account