Security at ModelTrack

How we protect your data and API keys. ModelTrack is designed so that sensitive information never persists — your keys pass through, your prompts stay private, and only cost metadata is stored.

1

API Key Handling

Your LLM API keys are passed through to providers in real-time. The proxy reads the Authorization header, forwards the request, and returns the response. That's it.

  • ✓Keys are NEVER stored, logged, or persisted
  • ✓Keys exist only in memory during the request (typically < 1 second)
  • ✓Keys are never sent to any ModelTrack service or third party
  • ✓The proxy never modifies your Authorization header

Data flow

Your App
  │
  │  Request + API key (Authorization header)
  ▼
ModelTrack Proxy (in-memory only, < 1 second)
  │
  │  Forwards request + API key to provider
  ▼
LLM Provider (Anthropic, OpenAI, etc.)
  │
  │  Response
  ▼
ModelTrack Proxy
  │
  ├──▶ Cost event (model, tokens, cost, latency) → Database
  │    (NO key, NO prompt, NO response content)
  │
  │  Response (unmodified)
  ▼
Your App
2

What We Track

ModelTrack records only the metadata needed for cost tracking and attribution. Here is exactly what is and is not included in a cost event.

What IS tracked

  • ✓Model name
  • ✓Input & output tokens
  • ✓Cost (USD)
  • ✓Latency (ms)
  • ✓Provider
  • ✓Team / App / Feature headers
  • ✓Timestamp
  • ✓Cache hit status
  • ✓Request ID

What is NOT tracked

  • ✗Prompt content
  • ✗Response content
  • ✗API keys
  • ✗Authorization headers
  • ✗Request/response bodies

Example cost event

cost event (JSON)
{
  "timestamp": "2026-03-26T10:30:00Z",
  "request_id": "req_abc123",
  "provider": "anthropic",
  "model": "claude-sonnet-4-6",
  "input_tokens": 150,
  "output_tokens": 320,
  "cost_usd": 0.00234,
  "latency_ms": 1200,
  "team": "ml-research",
  "app": "chatbot",
  "feature": "customer-support",
  "session_id": "user-query-123",
  "cache_hit": false,
  "routed": false
}
3

Infrastructure

Our hosted service runs on Google Cloud with security best practices:

  • ✓

    Google Cloud Run

    Serverless, auto-scaling containers. No persistent disk — the proxy is completely stateless.

  • ✓

    Google Firestore

    Cost events and account data are stored in Firestore with encryption at rest (AES-256).

  • ✓

    SSL/TLS everywhere

    All connections — from your app to the proxy, and from the proxy to LLM providers — use TLS encryption.

  • ✓

    Stateless proxy

    The proxy holds nothing in persistent storage. Each request is independent. There is no disk, no local database, no file system state.

4

Self-Hosted Option

ModelTrack is open source. If you need maximum control over your data, run the entire stack in your own infrastructure.

  • →Deploy the proxy, API, and dashboard in your own VPC
  • →Your data never leaves your infrastructure
  • →Full control over networking, access, and encryption
  • →Same features as the hosted version

Get started with self-hosting: github.com/ModelTrack/modeltrack

5

Reporting Issues

We take security seriously. If you discover a vulnerability or security concern, please let us know.

  • →Email: security@modeltrack.ai
  • →Responsible disclosure is welcome and appreciated
  • →We will acknowledge receipt within 48 hours
  • →We will not take legal action against good-faith security researchers

For general support questions, reach out to support@modeltrack.ai.