@securitychecks/mcp

Security verification for AI-assisted development — MCP server for backend security invariant checks.

MCP server that lets Claude verify security invariants in your code — auth enforcement, injection safety, access control, webhook idempotency.

What is this?

Your AI assistant writes code. This gives it the ability to review that code for production-readiness.

The loop: Copilot/Cursor writes → Claude reviews via MCP → Ship with confidence.

Claude can now check for the patterns that cause production incidents, based on what staff engineers actually catch in review.

Installation

npm install -g @securitychecks/mcp

Usage with Claude Code

Add to your Claude Code MCP configuration:

{
  "mcpServers": {
    "scheck": {
      "command": "scheck-mcp",
      "args": [],
      "env": {
        "SCHECK_MCP_ALLOWED_ROOTS": "."
      }
    }
  }
}

Allowed roots (required)

For safety, scheck-mcp will only run inside the allowed roots. If you don’t set SCHECK_MCP_ALLOWED_ROOTS and the server is not started inside a git repository, it will refuse to scan.

Available Tools

scheck_run

Run scheck on the codebase.

Arguments:
- path (optional): Target path to audit
- include_context (optional): Include code context snippets in results
- max_findings (optional): Limit number of findings returned (default: 200)
- only (optional): Only run specific invariant checks by ID
- skip (optional): Skip specific invariant checks by ID

scheck_list_findings

List current findings from the last run.

Arguments:
- severity (optional): Filter by severity (P0, P1, P2)
- include_context (optional): Include code context snippets in results
- max_findings (optional): Limit number of findings returned (default: 200)

scheck_explain

Explain an invariant - what a staff engineer would say about it.

Arguments:
- invariant_id: The invariant to explain (e.g., "AUTHZ.SERVICE_LAYER.ENFORCED")

scheck_list_invariants

List all patterns a staff engineer checks for.

scheck_generate_test

Generate a test skeleton to prove an invariant is satisfied.

Arguments:
- invariant_id: The invariant to generate a test for
- framework (optional): Test framework (jest, vitest, playwright)
- context (optional): Extra context to generate a more targeted test

scheck_feedback

Report whether a finding was a true positive or false positive.

Arguments:
- invariant_id: Invariant ID (e.g., AUTHZ.SERVICE_LAYER.ENFORCED)
- verdict: true_positive or false_positive
- reason (optional): not_applicable, acceptable_risk, wrong_location, outdated_pattern, missing_context, other

Example Session

User: Check my code for issues a senior engineer would catch

Claude: [calls scheck_run]

Found 2 issues a staff engineer would flag:

1. **AUTHZ.SERVICE_LAYER.ENFORCED** (P0)
   Service "MembershipService" has exports without auth checks
   Location: src/services/membership.ts:12

   A staff engineer would ask: "What happens when a background
   job calls removeMember() directly, bypassing the route?"

2. **WEBHOOK.IDEMPOTENT** (P0)
   Webhook handler missing idempotency check
   Location: src/api/webhooks/stripe.ts:45

   A staff engineer would ask: "What happens when Stripe
   retries this webhook?"

User: Explain the webhook issue

Claude: [calls scheck_explain with invariant_id="WEBHOOK.IDEMPOTENT"]

Webhooks can be delivered multiple times. Without idempotency,
you might double-charge customers, send duplicate emails, or
corrupt data...

Why MCP?

AI writes code fast but doesn't reason about production scenarios:

• Webhook retries → double-charges
• Internal service calls → auth bypass
• Transaction rollbacks → phantom emails

This MCP server gives Claude the ability to catch these patterns — the things AI-generated code routinely misses.

Enterprise

For teams with compliance requirements:

• Audit trails: Every AI-assisted review is logged
• Local analysis: SOC2 compliant — no source code transmission
• Consistent patterns: Same staff check for all developers

License

Apache-2.0. See LICENSE for details.