AI APIs
Monitoring for AI APIs and inference endpoints
AI APIs
Why AI endpoints fail in ways a ping won't catch
- An upstream model provider has an outage, or one region is degraded.
- A 200 response carries an error payload or empty output.
- Cold starts and queueing make responses slow, not dead.
- An expired key or a quota limit returns 401 or 429.
Checks
What Pingara checks
HTTP with method, headers and body
Custom method, headers and request body on every plan. Call your own health route, or a small inference request.Expected status codes
Choose the codes that count as up, for example only 200, so 401, 429 and 5xx fail the check.Keyword on the response
The response must contain a string you choose, or must NOT contain one, such as "error". It is a substring check.Timing and Degraded
Each check records DNS, connect, TLS, time to first byte (TTFB) and total time. A monitor can be marked Degraded after repeated slow responses.WebSocket and TCP
A WebSocket monitor checks the handshake, and can send a message and expect a reply. TCP covers a streaming or self-hosted server.Request size
Headers up to 4 KB, body up to 8 KB, so keep a test prompt small.Regions and how often
All 4 regions on Pro, 1 region on Free. Every 30 seconds on Pro, every 5 minutes on Free. US East (N. Virginia), US West (Oregon), Europe (Ireland) and Asia Pacific (Singapore).Status page and summaries
A public status page, and AI-written incident summaries on resolved incidents.
Examples
Example monitor setups
Your inference service health
GET https://inference.example.com/health. Expected status 200. Keyword must contain "status":"ok". Mark the monitor down. All 4 regions on Pro. Every 30 seconds on Pro, every 5 minutes on Free.
A tiny real request
POST https://api.example.com/v1/generate. Header "Authorization: Bearer <key>". A small body. Keyword must contain "output", or the field you expect. Mark the monitor degraded. A paid call costs a request on every check, so use an interval of 5 to 10 minutes. Everyone in your organisation can see the header value.
A streaming endpoint
WebSocket monitor on wss://api.example.com/stream. Optional message to send, and text the reply must contain.
A self-hosted GPU box
TCP monitor on gpu.example.com port 8000.
A status page
Add the monitors to a public status page so users can see whether the model API is up.
FAQ
Questions
Can I monitor an API that needs an API key?
Yes, add it as a custom header. Header values are stored with the monitor settings and every member of your organisation can see them, so use a restricted key with a low spend limit.
Will monitoring a model API cost me tokens?
Only if you point it at a paid inference call; a health endpoint avoids that.
Can Pingara detect an error hidden in a 200 response?
Yes, with a "must NOT contain" keyword check.
How fast will I know about an outage?
About a minute at 30-second checks on Pro (2 consecutive failures per region, then quorum).
Do you monitor OpenAI/Anthropic status for me?
You can point a monitor at any public HTTP endpoint you're allowed to call; Pingara doesn't ship pre-built provider monitors.
What does it cost?
Free for one monitor; Pro $9/month.
AI APIs