Metering

Metering

Metering is the aggregation of tokens, cost, and call counts across every governed call.

Per-session totals roll up to per-agent → per-project → per-org. Operators answer "what did this cost?" and "who's about to blow the budget?" from one place.

Metering is derived, not stored separately — it's a query (and pre-aggregated rollups) over the audit trail's call events.

At a glance

Derived from	Audit trail events on every LLM and tool call
Tracks	input tokens, output tokens, cost (USD), call count, latency p50/p99
Rolled up by	session → agent → project → org · day / week / month
Drives	cost dashboards, budget alerts, the cost.budget.* pipeline-step family
Status	next

What it tracks per call

Every LLM and tool call writes a row carrying:

Field	Source
input_tokens	LLM provider response usage block
output_tokens	LLM provider response usage block (streamed)
cost_usd	input/output × per-vendor unit price (current pricing table)
latency_ms	wall clock from request entry to response complete
tool_call_count	tool/MCP calls during the same session
cache_read_tokens / cache_write_tokens	for vendors that support prompt caching

How rollups work

Pre-aggregated counters in PostgreSQL keep dashboard queries fast:

       Session totals (1 row / session, finalized at session end)
            │
            ▼
       Daily agent totals (1 row / agent / day)
            │
            ▼
       Daily project totals (1 row / project / day)
            │
            ▼
       Daily org totals (1 row / org / day)

Live (in-flight session) counters live in the Session's context — finalized on session end.

Why this concept exists

Metering is the economic view of a session, distinct from:

• Pipeline findings — what was caught
• Tool footprint — what was done
• Metering — what it cost + how much was used

Metering also drives enforcement: the cost.budget.session, cost.budget.agent, cost.budget.org pipeline steps read live counters and reject calls when limits are hit.

Surfaces

Persona	Surface	What they see
Operator	Costs dashboard	Daily/weekly burn per project, top-N agents, per-vendor split
Operator	Per-agent detail	Live session burn vs budget, history
Finance	Monthly statement	Org rollup, broken down by team / project
Pipeline step	cost.budget.* reads counters	Reject calls when threshold exceeded
BYOK customer	BYOK dashboard	Same data scoped to their own credentials

Related concepts

• derived from ← Audit trail
• scoped to ← Session → agent → project → org
• drives → cost.budget.* pipeline steps (enforcement)
• adjacent to ↔ Pipeline findings, Tool footprint