How do I authenticate the scrape?

With a Cognocient key as a Bearer token. A dedicated proxy key is a good fit: it can read analytics but cannot manage your team, billing, SSO or provider credentials.

Are these counters or gauges, and what happens on the 1st of the month?

Spend, requests, tokens and latency are month-to-date counters that reset at the start of each UTC month. Prometheus handles counter resets in rate() and increase(), so alerts keep working across the boundary.

Does it include prompt or response content?

No. Only aggregate spend, counts, tokens and latency.

How do I scrape Cognocient spend into Prometheus or Grafana?

Q: Will a large number of feature tags overwhelm my Prometheus?

No. Series are capped at the 500 highest-spend provider, model and feature combinations; the rest are folded into feature="__other__".

A Prometheus metrics endpoint for month-to-date AI spend, requests, tokens, latency and budget utilisation, scrapeable into Prometheus, Grafana Cloud or any compatible tool. Growth plan and above.

GET /api/metrics/prometheus returns your AI spend in the Prometheus text exposition format, so you can graph and alert on it next to the rest of your infrastructure. Available on the Growth plan and above.

Scrape configuration

scrape_configs:
  - job_name: cognocient
    metrics_path: /api/metrics/prometheus
    scheme: https
    scrape_interval: 60s
    authorization:
      credentials: "<your Cognocient key>"
    static_configs:
      - targets: ["api.cognocient.com"]

Metrics

Metric	Type	Labels	Meaning
`cognocient_spend_usd_total`	counter	provider, model, feature	Month-to-date spend in USD
`cognocient_requests_total`	counter	provider, model, feature, status	Requests by HTTP status class (`2xx`, `4xx`, `5xx`)
`cognocient_tokens_total`	counter	provider, model, feature, direction	Tokens, `direction` is `prompt` or `completion`
`cognocient_request_latency_ms_sum`	counter	provider, model, feature, status	Total latency in ms; divide by requests for the mean
`cognocient_budget_used_ratio`	gauge	budget, scope, mode	Budget spend divided by its monthly limit
`cognocient_budget_limit_usd`	gauge	budget, scope, mode	Configured monthly limit

Untagged calls appear with feature="unattributed".

Example alerts

# Spend rate: more than $50 in the last hour
increase(cognocient_spend_usd_total[1h]) > 50

# Any budget above 90% used
cognocient_budget_used_ratio > 0.9

# Error rate above 5% for a model
sum by (model) (increase(cognocient_requests_total{status="5xx"}[15m]))
  / sum by (model) (increase(cognocient_requests_total[15m])) > 0.05

Month-to-date, and bounded

Counters reset at the start of each UTC month, which Prometheus handles in rate() and increase(). Series are capped at the 500 highest-spend provider, model and feature combinations; the remainder is folded into feature="__other__", so a tenant with thousands of tags cannot overwhelm a scraper.

How do I scrape Cognocient spend into Prometheus or Grafana?

Scrape configuration

Metrics

Example alerts

Frequently asked questions

On this page