Coding-agent control plane

Retained lifetime tokens

—

Only directly measured providers are included.

Codex Unknown; OpenCode Go Unknown; OpenRouter Unknown.

24 h model spend (tokens)

—

Prompt + completion across top measured models.

Metered cash spend

Unknown

Provider-charged money only (OpenRouter prepaid credits, direct API). Subscription-included usage is valued separately and is never counted here. Unknown

Allowance runway

Unknown

Codex-emitted allowance and reset status; credit balance stays private.

How to read this: Green means the telemetry route is working. Red means a component is down or an export is failing. The Collector receives spans once and converts them locally into the metrics this site reads, so nothing is shipped to an outside trace service. Values stay Unknown unless directly measured; Unknown never means zero.

Inference-source budget matrix

Lifetime tokens per provider; subscription, overflow and refill/credit fields stay Unknown unless the source publishes them.

lifetime

Subscription vs overflow

Split is shown only where a source distinguishes plan-billed from metered usage; otherwise it stays Unknown.

Cash vs subscription-equivalent

Metered rows are money charged per token. Notional rows are subscription-included usage valued at published per-token prices, shown for comparison only. Values stay Unknown when no authoritative source reports them.

Who bills for what

Every measured dollar, attributed to a named bill. The rows add up to the total, and only metered is money actually spent; subscription-included usage is valued at list prices for comparison. A dash means that bill has no cost on that basis. The final row is Claude Code sessions that record no provider, so no bill can be identified.

What the two token totals mean

Two different measurements, not two answers to one question.

OpenRouter cost authority

Account, key and session attribution are three different scopes and are never summed. The unattributed gap is routed traffic that recorded no native cost in its session store.

Context & cache efficiency

Lifetime prompt tokens per output token by agent, with direct cache read/write counters where available. Cache share stays Unknown for harnesses that do not export those counters.

lifetime

Source × harness

Lifetime prompt tokens grouped by provider and agent harness.

Attribution quality & freshness

How much measured traffic carries a usable model/provider label, and packet freshness.

Tokens by agent harness

Lifetime prompt and completion totals; the session count includes only records that report token usage.

lifetime totals

Top models

Model identifiers are public; counts are aggregates only.

24 h window

Top projects

Projects appear only as one-way hashes; original names are never published.

24 h window

Retained lifetime tokens

Provider-side totals kept in local source-of-truth databases.

lifetime totals

Coverage & freshness

What the exporter can and cannot currently see. Unknown is honest, not zero.

—

Sources

Coverage

Public token history

One aggregate packet per hour, retained for up to 90 days.

Starting

What is public—and what is not

This is a read-only mirror of the useful Grafana panels. Your browser downloads one small aggregate packet from a public GitHub data repository; it never connects to the private Grafana, Prometheus or Mac services.

Published intentionally: aggregate token and cost totals, provider and agent categories, model identifiers, and one-way project hashes. These are counts, not content.

Never published: prompts, responses, code, reasoning, tools, credentials, task names, paths, or task/session/trace identifiers.

—

—Failed exports · 1 h

—Received · 1 h

—Metric points exported · 1 h

—Work active

—Alerts firing