Telemetry

What We Collect

Every request generates this exact payload:

1{
2  "device_id": "anon_8f3a2b1c...",
3  "task_type": "code_review",
4  "model": "claude-sonnet-4",
5  "tokens_in": 1847,
6  "tokens_out": 423,
7  "latency_ms": 2341,
8  "success": true,
9  "cost_usd": 0.02
10}

What We NEVER Collect

• ❌ Your prompts
• ❌ Model responses
• ❌ File paths or contents
• ❌ API keys
• ❌ Your identity or location
• ❌ Anything that could identify you or your project

Task Type Inference

Task types are inferred from request characteristics (token counts, ratios), never from prompt content:

• quick_task, Short input/output (< 500 tokens)
• code_review, Medium-long input, medium output
• generation, High output/input ratio
• classification, Low output/input ratio
• long_context, Input > 10,000 tokens
• content_generation, Output > 1,000 tokens
• tool_use, Request includes tool calls
• general, Default classification

Local Breakdown API

The local dashboard ships with a JSON breakdown endpoint for cost rollups by model or agent:

1# Cost rows grouped by model (last 24h)
2curl http://localhost:4100/v1/telemetry/breakdown?dimension=model&window=24h
3
4# Cost rows grouped by agent (last 7d), labels resolved from local agent registry
5curl http://localhost:4100/v1/telemetry/breakdown?dimension=agent&window=7d

Windows: 1h, 24h, 7d, 30d. Unknown windows fall back to 24h. All data is local, nothing leaves your machine.

Verify What's Sent

You can audit telemetry before it's sent:

1# See payloads before they're sent
2relayplane --audit
3
4# Run fully offline (no telemetry at all)
5relayplane --offline

Opt Out

1# Disable telemetry
2relayplane telemetry off
3
4# Check status
5relayplane telemetry status
6
7# Re-enable
8relayplane telemetry on

Live Telemetry API (local proxy)

The local proxy exposes a read-only telemetry API the dashboard polls. Endpoints are bound to localhost and never leave your machine.

GET /v1/telemetry/stats?days=N

Aggregate stats for the last N days. Includes a today summary with numeric fields, plus top-level latencyP95 and cacheHitRate for the window.

1{
2  "today": {
3    "totalCost": 1.23,
4    "totalRequests": 412,
5    "avgLatencyMs": 814,
6    "latencyP95": 1932,
7    "errorRate": 0.012,
8    "cacheHitRate": 0.41
9  },
10  "latencyP95": 1932,
11  "cacheHitRate": 0.41,
12  "summary": { "...legacy fields...": true }
13}

GET /v1/telemetry/spend-by-hour?day=YYYY-MM-DD

24 hourly buckets of spend for a single UTC day. Always returns 24 entries, zero-filled. Malformed day returns 400.

1[
2  { "hour": 0, "usd": 0 },
3  { "hour": 1, "usd": 0.04 },
4  { "hour": 2, "usd": 0.11 },
5  "...",
6  { "hour": 23, "usd": 0 }
7]

GET /v1/telemetry/providers?days=N

Per-provider rollup over the last N days. share is 0..1 and sums to ~1.0 across providers. health is 0..100.

1[
2  {
3    "provider": "anthropic",
4    "share": 0.62,
5    "p95Ms": 1840,
6    "rpm": 42,
7    "health": 98,
8    "primary": true,
9    "note": ""
10  },
11  {
12    "provider": "openai",
13    "share": 0.38,
14    "p95Ms": 920,
15    "rpm": 25,
16    "health": 100,
17    "primary": false,
18    "note": ""
19  }
20]

GET /v1/telemetry/runs

Each record now includes routing_rule (string or null) and routing_reason (string or null) so dashboards can surface why a request landed on a given model.

GET /v1/sessions

Each session record now includes total_savings_usd (number), the per-session sum of routing savings vs the all-Opus baseline.

GET /control/budget

Single source of truth for the dashboard budget meter. Shape: { limit_usd, today_usd, hourly_usd, this_week_usd, this_month_usd }. The dashboard derives the visible cap by multiplying limit_usd by the active TODAY/7D/30D window, preferring this_week_usd / this_month_usd when those are set.

Why We Collect This

Anonymous telemetry helps us:

• Improve routing decisions (which models work best for which task types)
• Track aggregate cost savings across the network
• Identify performance issues

Every user (free or paid) contributes to making routing smarter for everyone.