ArgusArgus

Find what's burning your API credits

A runaway loop, a forgotten cron, a retry storm: how to find the call that's draining a balance, from the burn rate down to the endpoint.

Start from burn rate, not the bill. Find the provider whose daily burn jumped against its 7-day average, find the hour it started from the balance series, then group per-call records by endpoint, model and status to find the call. Tag your calls with a job name and the leak points at code.

The usual suspects

LeakWhat it looks like
Retry loop on a non-transient errorThousands of identical calls in minutes. Worst when the provider's out-of-credit error looks retryable: OpenAI returns 429, Perplexity 401.
A cron that ran wider than intendedA step in the balance at the same hour every day. A backfill that never got its date filter.
Agent loop with a growing contextInput tokens per call climb run over run while output stays flat. Cost per call doubles, then doubles again.
Prompt cache missesCached tokens drop to zero after a deploy that changed the prompt prefix. Same traffic, higher bill.
Crawl depthFirecrawl charges about a credit per page, so a crawl without a page limit spends in proportion to the site, not your intent.
Staging on the production keySpend with no matching production traffic. Only visible if calls carry an environment tag.

Narrow it down

  1. 1
    Find the provider

    Today's burn over the 7-day average. Argus raises spend_spike when a provider is up 50% or more week over week, and credits_low when the new rate leaves seven days or less.

  2. 2
    Find the hour

    A balance polled hourly gives you a time series even for providers with no usage API. A step at 02:00 every night is a schedule. A ramp that started at 15:40 is a deploy or something you ran.

  3. 3
    Find the call

    Group per-call records by endpoint, model and status. Look for paid calls that returned errors, the same endpoint called with the same parameters, and input tokens that grow per call.

  4. 4
    Find the caller

    Endpoints don't tell you which job made the call. Put that in metadata on each event (string key/values, up to 2 KB):

    {
  "id": "01JABZ6Q4V8R3M",
  "hostname": "api.firecrawl.dev",
  "endpoint": "/v2/scrape",
  "method": "POST",
  "status": 200,
  "latency_ms": 2310,
  "units": { "requests": 1 },
  "metadata": { "job": "nightly-enrich", "env": "staging" }
}
  5. 5
    Cap it

    One key per service, each with its own limit. OpenRouter keys take a credit limit (visible as limit and limit_remaining on /api/v1/key). OpenAI supports spend limits per project. Anthropic supports spend limits per workspace.

Hand it to your agent

With the Argus MCP server connected, the whole hunt is one prompt:

Use Argus. Call get_advisories, then get_request_costs for any provider
with spend_spike or credits_low (limit 100). Group the calls by endpoint, model
and status, and tell me which one accounts for most of the cost and when it
started.
  • Per-call records need the SDK or ingest API. Polling alone gives you the provider and the hour, not the call.
  • Stop a leak at the source: don't retry 402s, out-of-credit 429s or Perplexity's 401. Every provider's error at zero.

Questions

How do I find which API key is using my OpenRouter credits?
GET https://openrouter.ai/api/v1/key with each key returns that key's usage and limit. Give every service its own key with its own limit and the leaking one identifies itself.
Why did my credits drain overnight?
The usual causes are a scheduled job that ran more than intended, a retry loop on an error that isn't transient (such as an out-of-credit 429 or 401), or an agent loop resending a growing context. The hourly balance series shows which: a step points at a schedule, a steady ramp at a loop.

Related

OpenRouter API usageFirecrawl API usageTavily API usage
Let your coding agent track API spend with MCP

Give Claude Code, Cursor or Codex a read-only view of every metered API you build with, so it can check what a run cost, spot a leak, and notice a provider running dry.

What happens when API credits run out

The status code and error each provider returns at zero balance. Several look like rate limits or bad keys, so they get retried instead of fixed.

Let your coding agent watch the meter.

Point your coding agent at Argus and ask which endpoint is spending the most. It reads the same ledger you would.

Start freeRead the MCP docs