Secure your app. Protect your revenue.

You don’t have an AppSec team. You have a product that bills.

A scanner dumps 200 lines. Claude sometimes refuses to audit. The hole that costs you is not a missing header: it is an IDOR between two accounts, a tool that is too open, a session that leaks. The kit runs on your machine. Karukera sees nothing.

You choose how to look.

14  const row = await find(id)
15  return json(row)
16
     ↑ no tenant check

How it works

The flow is short.

1. 1

   You clone the repo

   In Claude, Codex, Cursor or Hermes. On your machine.

2. 2

   You pick the project and the depth

   Express, Web, SaaS, MCP, delta, red team… Only a system you are allowed to audit.

3. 3

   You give access, or you don’t

   Test accounts to look from the inside. Otherwise the audit stays outside.

4. 4

   You get the report

   Evidence, statuses, and fix tickets to hand to your LLM.

Eight kinds of audit.

1. 1

   Express

   First pass. Surface, pages, secrets. 30–45 min.

2. 2

   Full Web

   Site or app: pages, cookies, front, same-origin API.

3. 3

   Full SaaS

   Orgs, roles, billing, isolation between tenants.

4. 4

   Agents / MCP

   Tools, skills, a copilot wired to data.

5. 5

   Delta

   After fixes: what moved.

6. 6

   Continuous

   Periodic snapshot, release guardrail.

7. 7

   Red team

   Adversarial exercise. Written mandate required.

8. 8

   Board report

   Exec summary from a journal QA already signed.

Claude can refuse. OpenRouter won’t.

You can do everything in Claude or Codex. If the model refuses the audit, deposit an OpenRouter key. About €10 in credits for one audit. Frontier models do not carry the same filter.

Eight depths. Pick from the product.

Express for a first signal. Full Web for a site. Full SaaS for isolation. MCP if the value is the agent. Each mode has its guide.

• 01-express · 30–45 min

  Website security audit: Express mode, a first signal.

  When to open Express, what it covers, and why an Express at confidence 5 is a fault.

• 02-complet-web · 1 day

  Web pentest: Full mode, for a site or an app.

  Pipeline of a full web audit: what runs, what stays Untested without an account, and when to switch to SaaS.

• 03-complet-saas · 1–2 days

  SaaS audit: isolation first, not headers.

  Two-tenant rule, pipeline 00–11, multi-tenant specialist. This is not SSPM (locking down Slack): it is auditing the SaaS YOU build.

• 04-agents-mcp · 4–8 h

  MCP security: audit the tools, not just the UI.

  Who may call which tool, a tool + another tenant’s id, exposed prompts and .env, third-party marketplace.

• 05-delta · 2–6 h

  Re-audit: what moved after the fixes.

  Snapshot before any new collection, re-check of open findings, no classes outside the baseline.

• 06-continuous · 30–90 min / run

  Periodic guardrail: Continuous mode.

  Without a baseline, this is the wrong mode. OpenRouter key on every run. Short QA only if a Confirmed is born.

• 07-redteam-leger · 1–3 days

  Light red team: active, authorized, no weapon.

  Four mandate clauses, what “light” allows, and why “it’s my site” in chat is not enough.

• 08-rapport-board · 1–3 h

  Board report: facts already held, nothing else.

  Hard gate: qa.passed=true, sign-off present, no finding newer than QA. Otherwise stop.

Before you clone.

Do I need to code?

No. You clone the repo and open it in Claude, Codex, Cursor or Hermes. Prompts, configs, templates.

Do I have to use OpenRouter?

No. You can stay in Claude or Codex. If the model refuses the audit, an OpenRouter key unlocks the ones that will. About €10 in credits for one audit.

Is this an online scanner?

No. The kit runs on your side. Karukera never sees your target or your code.

Can I use it on any site?

Only a system you have written authorization to audit.

What do I get?

A prioritized report, the evidence, six measurement statuses, and tickets with the prompt to paste into your LLM to fix.

Are updates paid?

No. Public repo, MIT license. git pull.