Prompt and model changes break quietly: no errors, no latency spike, just worse answers that users notice before you do. Repath brings the canary rollout to LLM features.

Point your existing OpenAI, Anthropic, Gemini or OpenRouter SDK at Repath (a base URL and one header). Define a rollout: baseline vs candidate prompt or model. Repath sends a small slice of live traffic to the candidate, an LLM judge scores sampled responses from both sides on the same criteria, off the request path, and every 30 seconds a controller advances the rollout (5% → 25% → 100%), holds, or rolls it back to 0%. Every decision is logged with its scores and the worst answers behind it, including the judge's reasons.

Useful right now for forced model migrations, like gpt-4o and gpt-4.1 leaving OpenAI's API on January 14, 2027, or Claude Sonnet 4.5 retiring on November 30, 2026.

The gateway adds about 0.4 ms at the median. 7-day free trial, no card; paid plans from $20/month.