“OpenAI is up” says nothing about your key, your model or your rate limit. PylonMon watches the model API your app actually depends on — and it does it without ever spending a token of your credit.
Works with OpenAI, Anthropic, Azure OpenAI, OpenRouter, vLLM, Ollama — anything with a models endpoint
# one free metadata call per check — four answers, zero tokens: Support bot — OpenAI — key ok · gpt-4o-mini listed · 412ms · headroom 82% # and the pages nobody else sends: Support bot: API key rejected (HTTP 401) — revoked, expired, or wrong key Summarizer: model "claude-3-haiku-20240307" is no longer listed — deprecated or renamed? Support bot: rate-limit headroom 9% — requests start failing when it hits zero
A revoked or expired API key is a full outage the provider's status page will never show. We authenticate with your key on every check and page you the moment it's rejected.
Providers deprecate and rename model snapshots constantly. If the model your code names disappears from the models list, you're paged before your users meet the error.
Real measured latency to the real API endpoint — with a slow-response threshold, graphs, and the same escalation ladders as every PylonMon monitor.
The provider reports your remaining rate limit on every response. We read it and warn you at the threshold you pick — before requests start bouncing off the limit.
Every check is a single GET of the provider's models endpoint — the same free call
their own docs use to list what your key can reach. It proves the API answers, authenticates your key,
confirms your model exists, measures latency, and carries the rate-limit headers. PylonMon never sends a
prompt, never requests a completion, and never spends a cent of your API credit. Your key is stored for
this and used for nothing else.
Want to try it with no key at all? Point an AI monitor at
https://pylonmon.com/aimock — our demo models endpoint (watch model
pylon-demo-1; a key containing “revoked” demos the failure path).
The model API is only half the dependency. The cron job that runs your nightly embeddings, the agent that triages your inbox — when they die, nothing errors; they just stop. A heartbeat monitor catches exactly that: the job curls a URL when it runs, and silence pages you. Ship the logs to Logs and a log-match monitor catches the 429s and context-length errors in between.
Anything OpenAI-compatible (OpenAI, Azure OpenAI, OpenRouter, Together, vLLM, LM Studio …), Anthropic, and Ollama for self-hosted models. Pick the provider, paste the base URL if it isn't the default, name the model.
It's used for the free models endpoint only — a read-only call that can't generate anything, can't spend anything, and can't see your data. You can rotate it any time; the monitor takes the new key in one edit.
AI monitors are ordinary monitors — they come with every plan, including free, and count like any other monitor. See pricing →
Free plan. No credit card. Zero tokens spent — ever.