Scan MCP Servers for Security Threats
Audit MCP servers for malicious tool definitions before installing
MCP servers can contain hidden instructions in tool descriptions that manipulate AI behavior. Scan them before adding to Claude.
Quick Start
# Scan an MCP server repo before installing
superagent scan --repo https://github.com/user/mcp-serverWhat It Detects
- Hidden instructions in tool descriptions
- Prompt injection in README/docs
- Malicious default configurations
- Data exfiltration patterns in tool implementations
Example Workflow
# 1. Find an MCP server you want to use
MCP_REPO="https://github.com/someone/cool-mcp-server"
# 2. Scan it first
superagent scan --repo $MCP_REPO
# 3. Review the report, then install if clean
claude mcp add cool-server -- npx -y cool-mcp-serverProgrammatic Usage
import { createClient } from "safety-agent";
const client = createClient();
async function auditMcpServer(repoUrl: string) {
const { result } = await client.scan({ repo: repoUrl });
return result;
}The result is a comprehensive natural language report that can be passed directly to a downstream model for interpretation or decision-making.
Environment Variables
SUPERAGENT_API_KEY- RequiredDAYTONA_API_KEY- Required for sandbox execution