Skip to main content

Monitor your project

What you can watch, and where​

You want to seeConsoleAPI, with a virtual key
The spend of today, the last week or the last monthOverview › Spent, Spend over timeGET /usage/me: the spend of this month
The spend of each app, key and modelOverview › Per-app burn, Per-key burn, Model splitGET /usage/me: by_model, by_task
The calls that failed, and whyPerformance › Failures: Failures by reason in words, and the log—
The vendors that fail oftenPerformance › Failures › Failures by provider, and Models › Models and vendorsGET /status/peek: the state of each vendor
What the cache saved, and who callsUsage › Cache, Who calls—
Which model is fast, and what a call costsPerformance › Health › Latency band, then Models › Speed against cost and Vendor scorecardGET /api/me/requests/performance
Which models get the most calls, and how each of them performsPerformance › Models › Vendor scorecard: the eight models with the most calls, then Show allGET /api/me/requests/performance: calls of each model, and total
How often a conversation reached its goal, for each modelPerformance › Models › Vendor scorecard, column Goal reachedGET /api/me/requests/performance: the goal of each model
The tokens, the cost and the calls over timeUsage › the chart, Tokens by kind, By modelGET /usage/series
The models and vendors that your project skips nowPerformance › Models › Models and vendors › Retired now—
If the next call will be refused—GET /status/peek: tenant_capped
What to change to spend lessImprove › What to change—

The MCP tools proxium_read_performance, proxium_read_usage_series and proxium_list_requests answer the same numbers as the routes in the last column.

GET /api/me/requests/performance reads the report of each model over every project of the key. The answer holds every model of the window and total, which says how many models it holds. The report of one project is at GET /api/teams/{slug}/requests/performance, and the console reads it there with your session. The cost of a model is cost_usd in United States dollars and cost_millicents in millicents: one millicent is one hundred thousandth of a dollar. answered is the number of answered calls.

The console pages are explained in Console screens.

Get an alert​

Proxium sends an alert to Slack, Discord, Telegram or a webhook when the spend of the project or of one app passes a level. See Get spend alerts.

Proxium also sends an alert to the same channels when your failover policy retires a model or a vendor. The alert is on by default. To turn it off, clear Send an alert in Routing › When a call fails.

Send call records to your own tools​

Proxium sends a record of each call to your webhook or OpenTelemetry backend. See Send call records to your tools.

Watch the spend from your own system​

To alert from your own system, read the spend on a schedule.

Example: a script that a scheduler runs each hour. It exits with an error when the spend of the month passes a level that you choose.

check-spend.sh
#!/bin/sh
used=$(curl -s https://proxium.tech/usage/me \
-H "Authorization: Bearer $PROXIUM_KEY" | jq '.credits.used_usd')
echo "Proxium spend this month: \$$used"
awk -v u="$used" 'BEGIN { exit (u > 50) ? 1 : 0 }'

To stop an app at a limit, do not wait for an alert. Set a limit, and Proxium refuses its calls above it. See Set budgets and limits.

Watch a self-hosted gateway​

A self-hosted gateway serves /metrics in the Prometheus text format. The page carries the state of each vendor, the failover outcomes and the latency of the gateway itself:

SeriesWhat it shows
axon_circuit_state{provider,state}The state of the circuit breaker of each vendor: 1 on the state that holds, 0 on the other two
axon_failover_total{outcome}The calls through the failover chain: first_hop, later_hop or exhausted
axon_overhead_seconds{source}The wall time of a call minus the time of its upstream attempts
axon_classifier_seconds{by}The wait of the classifier, by who chose the tier
axon_cache_seconds{kind}The wait on the response cache, by question, exact or store
axon_redis_seconds, axon_postgres_seconds{op}The wait on Redis, and on a Postgres read of the request path

The label source of axon_overhead_seconds holds the caller: chat, pi, memory, console, bot or other. The label by of axon_classifier_seconds holds who picked the tier: system_one, llm, fallback, timeout, reused or other. The label op of axon_postgres_seconds holds the read: key_lookup, catalog, pricing, settings or routing.

When axon_overhead_seconds rises and the time of the vendors does not, the gateway spends the time itself. Read the four dependency histograms of the same window to see where it waits.

Put the page behind a token. Set AXON_METRICS_TOKEN to a value that you choose. A scrape must then send it as Authorization: Bearer <token>, and a request without it gets a 401. Leave the variable unset when only a Prometheus inside the cluster can reach the Service. The page stays open, which is the default.

curl -sS -H "Authorization: Bearer $AXON_METRICS_TOKEN" https://your-gateway/metrics

dashboards/axon.json in the repository shows the two histograms of the gateway, the circuit state and the failover outcomes. Import it into Grafana and pick your Prometheus data source.

What is not available yet​

FeatureState today
Spend alerts by emailNot available. Use Slack, Discord, Telegram or a webhook