← Back to home

Architecture & Privacy

Last updated: April 25, 2026

This page documents how Zarev Pro processes your data technically. The backend code stays closed source but the architecture is public: you must be able to know exactly what leaves your machine.

Data flow

When you run the Zarev CLI, here's exactly what happens:

┌─────────────────────────────────────────────────────────────┐
│  YOUR MACHINE (local-first)                                 │
│                                                             │
│  ~/.claude/         scan + parse      ┌─────────────────┐   │
│  ├─ .mcp.json   ─────────────────────▶│  CLI (Node)     │   │
│  ├─ skills/                           │                 │   │
│  └─ hooks/                            │  Gitleaks strip │   │
│                                       │  (~150 patterns)│   │
│  ~/.zarev/                            │       │         │   │
│  └─ audit.log  ◀──── append ──────────│       ▼         │   │
│                                       │  Preview UI     │   │
│                                       └─────────┬───────┘   │
└──────────────────────────────────────────────────│──────────┘
                                                   │ HTTPS / TLS 1.3
                                                   │ Bearer token
                                                   ▼
┌─────────────────────────────────────────────────────────────┐
│  ZAREV BACKEND (Supabase EU + Vercel)                       │
│                                                             │
│  POST /api/v1/zarev/match  ─────▶ Supabase EU (Frankfurt)   │
│         │                            │                      │
│         │                            ▼                      │
│         │                         RLS scoring               │
│         │                                                   │
│  POST /api/v1/zarev/brief-prompt ─▶ Anthropic (US, SCC+DPF) │
│                                       Claude Haiku          │
│                                       │                     │
│                                       ▼                     │
│                                       JSON response         │
│                                       (5 min cache)         │
└─────────────────────────────────────────────────────────────┘

Threat model

The 4 main threats we address in V0.1:

1. Secret leak (API keys, tokens)

Mitigation — Gitleaks pre-strip with ~150 patterns (sk-*, GITHUB_TOKEN, AWS_*, OPENAI_API_KEY, etc.) + heuristic on long hex/base64 values in .env* files. Mandatory UI preview before each upload: you see the exact payload, you can cancel.

2. Network man-in-the-middle

Mitigation — TLS 1.3 mandatory, invalid certs rejected by the CLI. No HTTP fallback.

3. Prompt injection via brief

Mitigation — the Anthropic system prompt strictly isolates the stack context from user content. No exfiltration possible: the brief has no access to past conversations or other users' DB.

4. Data exfiltration via generated brief

Mitigation — the brief contains only Claude recommendations, never raw echo of the scanned stack. Output payload is validated server-side before returning to the CLI.

Fields sent to the backend

Exhaustive list of what the CLI sends to POST /api/v1/zarev/match and POST /api/v1/zarev/brief-prompt:

  • stack_hash: SHA-256 of stack fingerprint (deterministic, no PII).
  • mcps: list of detected MCPs (slug, version, transport — no config values).
  • skills: list of Skills (slug, version — no prompt content).
  • hooks: list of Claude Code hooks (event, script-name — no body).
  • runtime: os (darwin/linux/win32), node version, claude-cli version. No hostname or username.
  • user_id: Supabase UUID, linked to your Pro account. Sent only when authenticated.

Before/after strip example

Before strip (local .mcp.json file):

{
  "mcpServers": {
    "supabase": {
      "command": "npx",
      "args": ["-y", "@supabase/mcp-server-supabase"],
      "env": {
        "SUPABASE_URL": "https://xyz.supabase.co",
        "SUPABASE_SERVICE_ROLE_KEY": "eyJhbGc...REDACTED..."
      }
    }
  }
}

After strip (sent to backend):

{
  "stack_hash": "sha256:b1946ac9...",
  "mcps": [
    {
      "slug": "supabase",
      "transport": "stdio",
      "version": "0.1.4"
    }
  ],
  "runtime": {
    "os": "darwin",
    "node": "20.11.1",
    "claude_cli": "1.4.2"
  }
}

Going further