Try MCP security scanning in 30 seconds — `npx correctover-scan --demo`

Correctover is an AI agent runtime authorization and evidence verification layer. We just shipped v2.4.7 with a zero-install CLI demo.

Try it now

npx correctover-scan --demo

This scans five sample MCP servers — four intentionally vulnerable, one clean — and reports CRITICAL and HIGH findings with remediation:

• dangerous_command — piped remote scripts to bash, sudo escalation
• stdio_env_exposure — sensitive environment variables leaked via stdio transport
• plaintext_http — unencrypted MCP endpoints
• overprivileged_args — excessive tool arguments

What Correctover does

1. Before tool execution: blocks SSRF (cloud metadata, internal IPs), command injection, credential exfiltration
2. After tool execution: scans outputs for leaked secrets and prompt injection
3. Every decision gets an Ed25519-signed receipt for audit trails

The verification core runs at P50 ≈ 2.7μs in Node.js with zero production dependencies. It implements the CCS 7-dimension runtime verification standard, published as an IETF Internet-Draft.

5-minute integration

import { GuardrailProvider } from 'correctover/guardrail';

const guardrail = new GuardrailProvider({
  policies: { blockSSRF: true, blockInjection: true, blockCredentialLeak: true },
  enforce: false, // audit-first: observe and log, switch to true when ready
});

// Wrap any MCP tool call
const decision = await guardrail.evaluateToolCall({
  tool: 'execute_command',
  args: { command: 'curl http://169.254.169.254/latest/meta-data' }
});
// decision.allow === false (SSRF to cloud metadata endpoint)

Links

• npm: https://www.npmjs.com/package/correctover
• Website: https://unpkg.com/correctover-site@1.0.0/index.html
• Source (Codeberg): https://codeberg.org/correctover-labs/correctover
• IETF draft: https://datatracker.ietf.org/doc/draft-correctover-ccs/

Feedback on false positives from real MCP traffic is especially welcome.