A leaked Supabase service key
CriticalThe service key skips every row-level rule you wrote, so once it ships in the browser bundle anyone who opens devtools owns the whole database.
Your assistant draws your system live on a canvas, then circles what will break.
Arketo is a standard MCP server. Your assistant connects to it once, then draws, routes and reviews your system while you talk.
“Map my app, then find what breaks at 50x the traffic.”
In Claude Code
claude mcp add --transport http arketo "<your-mcp-url>" --header "Authorization: Bearer <your-key>"Your code never leaves your machine. Your own AI describes the system, and Arketo only ever stores the picture.
You can build an app in an afternoon now. What you cannot do is see the system you just built. The bottleneck, the missing auth check, the single point of failure: none of it shows up until load, an audit, or an outage forces it into the open.
Working features fast. Code that runs. Files you can read top to bottom, one at a time.
The shape of the whole. Files are not architecture, and the diagram in your head is rarely the one you actually shipped.
Every component and connection, laid out into frontend, backend, and data tiers and routed automatically as you describe it. Drag to rearrange and the layout keeps up.
“a React app, a Node API and Postgres”
Bottlenecks, security gaps, and scaling limits, marked by severity right on the diagram where they live.
The service-role key is in the browser bundle
No rate limit on sign-in
A full Markdown write-up with a phased scaling plan, ready to drop into your repo or a ticket.
An assistant writes code that works. Working and safe are not the same thing, and these four gaps turn up again and again in apps that shipped fast. Arketo goes looking for each of them by name.
The service key skips every row-level rule you wrote, so once it ships in the browser bundle anyone who opens devtools owns the whole database.
The only thing deciding what may be read or written is code running on someone else's machine, and they can edit it.
One script can sit on your login or password-reset endpoint all night, and nothing in the app is counting.
If the signature is never checked, anyone who finds the URL can post “payment succeeded” and get what paying customers get.
Wiz found 1 in 5 vibe-coding platforms exposed. Researchers pulled 400+ leaked secrets out of 5,600 vibe-coded apps. Bots find a leaked key within minutes.
Real architectures already mapped and reviewed: the components, the connections between them, and the findings that came out of each. Read one before you run your own.
ChatGPT will sketch a diagram. So will Mermaid or Eraser. Then it is a picture: frozen the moment it is made, redrawn from scratch every time the code moves.
| What it does | A chat diagram, or Mermaid | Arketo, driven over MCP |
|---|---|---|
| Remembers your system | Re-prompted from scratch | Kept and updated in place |
| Knows which paths carry load | No idea | Hot paths marked on the wire |
| Tells you what will break | If you ask, in prose | By severity, on the component |
Not a drawing of your architecture. A running model of it, that your own assistant keeps current.
Each one stands for something specific: component type, connection load, finding severity. We built the site the same way, so every colour you have seen so far is data, not decoration.
The assistant you already pay for does the thinking. We host the live canvas, the findings, and the report. No AI key sits on our side, so there is no token margin to pay.
Same plan on every cadence. No seats, no tiers, no usage surprises. Cancel anytime.
Security scanners for vibe-coded apps charge $29 to $49 a month. Arketo is $7, because your own model does the thinking and there is no token margin to pay.
Anything else, and a person answers. There is only one of us, so it is a real reply.
Ask a questionNo. Arketo never reads your repository and never holds an AI key. Your AI assistant connects with one command and describes the system to the canvas over MCP, the open standard assistants use to reach outside tools. The thinking happens on the model you already pay for.
Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Codex, and anything else that speaks MCP. Your assistant connects with one command, then drives the live canvas for you.
The full taste: your system drawn and routed automatically, findings surfaced by severity in place, and the Markdown report. No card needed to start.
Everything, with no seats, tiers, or usage surprises. It is one plan billed your way: $2.50 weekly, $7 monthly, or $69 yearly. Because no model key sits on our side, there is no token margin baked in. Cancel anytime.
Yes. Diagrams are yours by default; sharing is opt-in per project via a link you control and can revoke.
You already describe systems to it in plain language. Let it draw the one you built, and tell you where it bends.
Map my app, free