Usage Tracking

Telemetry Events

Every request through the proxy generates a telemetry event with the following schema:

1interface TelemetryEvent {
2  device_id: string;       // Anonymous random ID
3  task_type: string;       // Inferred from token patterns
4  model: string;           // Model used (e.g., "claude-sonnet-4-6")
5  tokens_in: number;       // Input tokens
6  tokens_out: number;      // Output tokens
7  latency_ms: number;      // Request latency
8  success: boolean;        // Whether request succeeded
9  cost_usd: number;        // Estimated cost in USD
10  timestamp: string;       // ISO 8601 timestamp
11  requested_model?: string;       // Original model before routing
12  cache_creation_tokens?: number; // Anthropic cache creation tokens
13  cache_read_tokens?: number;     // Anthropic cache read tokens
14}

Cost Estimation

Costs are calculated using per-model pricing tables. The proxy knows pricing for all supported models:

Anthropic Cache Token Handling

When Anthropic prompt caching is used, cost estimation accounts for the different pricing tiers:

• Cache creation tokens — 1.25× the base input price
• Cache read tokens — 0.1× the base input price (90% discount)
• Regular input tokens — standard input price

Local Storage

Telemetry events are stored locally in two formats:

• ~/.relayplane/telemetry.jsonl — Append-only JSONL file of all events
• ~/.relayplane/history.jsonl — Rolling request history (max 10,000 entries, 7-day retention)

Stats Collector

The proxy maintains a rolling 1-hour statistics window with:

• Total, proxied, and direct request counts
• Success/failure counts
• Latency percentiles (p50, p95, p99)
• Circuit breaker state transitions

Access stats via the /status endpoint or the dashboard.

Dashboard Endpoints

The proxy exposes telemetry data via REST endpoints:

• GET /v1/telemetry/stats?days=7 — Summary stats with model breakdown
• GET /v1/telemetry/runs?limit=50 — Recent request history
• GET /v1/telemetry/savings — Cost savings analysis
• GET /v1/telemetry/health — Per-provider health status