Tab · Insights

Telemetry

Know what every Claude Code session actually cost. A combined view of tokens, spend, and per-developer rollups across every tracked repo — captured locally, retained as long as you want, ready to invoice or split.

See it in motion

Where to find it

  • Localhost: http://localhost:4000/telemetry.html — defaults to combined mode across every tracked repo.
  • Hosted: The team rollup view at repoops.ai/team/activity (per-binding drill-down).
  • Keyboard: ⌘ K then tel.
  • API: GET /api/telemetry?repo=<id> · POST /api/telemetry-pull?repo=<id>

What it does for you

You'll know what AI cost this week.Every session's input, output, and cache tokens × per-model pricing. Filter by repo, developer, day, or model. The cost of AI work stops being a guess.
You'll know who burned what.Per-developer rollups give you a defensible split for contractor billing or internal chargeback. Combined-mode reveals the hot repos and hot developers.
You'll keep the long memory.Telemetry is local-first and lossless — retain it as long as you want, no rolling 30-day cap. Export to CSV any time.

Configure

Telemetry reads the local JSONL files that Claude Code writes. For production-LLM telemetry (apps you've shipped), wire the per-repo export endpoint and pull on demand.

  • CLAUDE_CODE_HOME — where the Claude Code session log lives. Defaults to the per-platform user home.
  • TELEMETRY_EXPORT_TOKEN_<ID> — auth token used when pulling prod telemetry from Vercel. One per tracked repo.
  • Settings → Telemetry privacy — metadata-only vs. content-redacted capture. The "preview what's shared" panel shows before-and-after.
  • Cross-developer mode: bind a desktop via the linking-code flow; the per-developer rollup appears once at least two devices are bound.

Use it well

  1. Keep Telemetry in combined mode for the morning glance.

    It'll tell you which repo your AI is currently expensive in. Scope to a single repo when something looks off (a one-day spike, a slow week).

  2. Use the by-day rollup to spot ramp / wind-down patterns.

    Sustained 3am usage = a session pattern worth reviewing. Activity heatmap surfaces this as a 7×24 grid.

  3. Switch to Billing outcomes for invoicing.

    Same telemetry source, but wrapped with proof-of-work line items and the per-client filter your invoice needs. Export CSV; paste into the invoice.

  4. Pull prod telemetry on demand, not on a schedule.

    The POST /api/telemetry-pull route is cheap; cache locally. A daily schedule risks rate-limits and adds noise.

Examples

✓
Monthly invoice (Andy)
Switch to per-client view → Acme → export CSV for May → paste into the invoice template → done. The proof-of-work column auto-populates from the prompt log.
⚠
When a sub-contractor's rollup looks low
That can mean two things: low prompt volume (they did the work in-IDE without AI) OR a missed binding (their desktop isn't bound yet). Activity ledger's integrity badge will tell you which.

Read more