Docs

Connect

If IA4AI is down

What each way of connecting does when IA4AI cannot be reached, and how to take the gateway out of the path in about 30 seconds.

IA4AI runs on one server today, with no failover. So the question to answer before you connect is which of your calls depend on it.

Gateway Observe-only (OpenTelemetry, telemetry, the extension)
How it connects Your tool's base URL points at https://ai.odux.io/gateway/… Your tool calls the provider directly and sends IA4AI a copy of its usage
In the live path? Yes: IA4AI forwards every call No
What you get Exact cost per call, tools, cache use, cheaper-model tests Totals and counts per workspace and agent
If IA4AI is down Calls fail fast with an error: they never reach the provider, so the agent stops Nothing changes for your tools: only the usage report is late or lost

If nothing may sit between you and your provider, use observe-only. It never blocks your work; you give up the exact cost of each call and the cheaper-model tests.

#Take the gateway out of the path

You can do this at any time, without waiting for us, by pointing the tool back at the provider:

  • An OpenAI-compatible tool (Cline, the OpenAI SDK, LangChain…): change the base URL from https://ai.odux.io/gateway/v1 to the provider's own, such as https://api.deepseek.com/v1 or https://api.openai.com/v1.
  • Claude Code: remove ANTHROPIC_BASE_URL, or set it back to https://api.anthropic.com.
  • Codex: switch model_provider back to the provider's own.

The tool then needs your own provider key in place of your IA4AI key, because the provider does not know the IA4AI key. Switch back the same way once IA4AI is reachable.

#What we do on our side

  • A deploy that fails its health check rolls back by itself; measured at 17 to 35 seconds.
  • Customer data sits on an encrypted volume whose key is never stored on the server. After a restart, the service stays down until we unlock the volume by hand. That is deliberate: a stolen disk or snapshot is unreadable.
  • A measured time to rebuild from backup on a new server is still to come. See Security.

Questions about availability for a pilot: samstern@odux.io.