Skip to main content
The fastest integration is the one your coding agent does for you. Omnia publishes a single canonical prompt that Claude Code, Cursor, or any coding agent with repository access can follow end to end: it inspects your repo, picks the right integration path, makes the edits, and runs the setup check to prove the result.
Paste the fetched prompt into your agent, provide OMNIA_API_KEY when asked, and read its report.

​
Integration paths

The prompt has your agent pick one of the four integration paths, based on what your repo actually uses: Gateway (your app calls an OpenAI-compatible API and can use Omnia’s catalog): points your existing client at https://gateway.omnia-voice.com/v1 with your Omnia key. Model names pass through unchanged. The agent also tags your traffic (X-Omnia-Tag) so it lands as one evaluation population, and threads a run id (X-Omnia-Trace-Id, or your existing W3C traceparent) when your app makes several calls per run. Your provider key (BYOK) (you run GPT, Claude, or Gemini and intend to keep it): same base-URL swap, but your model doesn’t change; you paste your provider’s API key into Settings → Workspace → Provider keys, the model name gains a provider prefix (anthropic/claude-sonnet-4-5), and your provider keeps billing you directly; Omnia bills nothing for those tokens. You get capture, tags, caching, and fallbacks with zero model risk. OpenTelemetry (your app already emits OTel traces: Vercel AI SDK telemetry, Pydantic AI, an existing OTel SDK): no packages, no code, three environment variables:
Omnia’s ingest understands the official gen_ai.* conventions plus the Traceloop, Vercel ai.*, and OpenInference attribute dialects, and normalizes them all into one gradeable shape. Tracing SDK (no OTel yet, and your inference stays where it is): one package, one line, standard OpenTelemetry underneath.
Pure-ESM Node apps skip even that line: node --import @omnia-voice/tracing/register. OpenAI, Anthropic, Gemini (Python), and LangChain are captured automatically, only libraries actually installed are instrumented. The SDK contains no instrumentation code of its own, and its README documents the identical setup in vanilla OTel: eject anytime, your spans don’t change.
Whichever path the agent picks, the traffic becomes gradeable population: you can grade it, calibrate a judge against your grades, and run screening comparisons; the BYOK, OpenTelemetry, and Tracing SDK paths do all of that before a single inference decision changes. Model-call content is stored only if your workspace has request logging enabled, under your retention window; span structure is kept either way.

​
Verifying by hand

The same check the agent runs:
It proves the key works, reports logging/traffic/grades/judge progress, and names the one next step. It installs nothing and never stores your key. Add -- --no-inference to skip the one-cent test completion.