Raycast v2.2 adds multi-provider AI routing, sparking debate on fair local vs hosted model evals

daani_maas · reddit · 2026-09-10

Raycast v2.2 lets Pro users route AI workflows through OpenAI-compatible providers, Ollama, or OpenRouter — but switching providers also changes tool support, retries, latency, and help-seeking frequency, complicating comparisons.

The author proposes freezing tool schema, permissions, input fixtures, and acceptance checks, then comparing task success on hidden checks, total cost, tool-call counts, recovery after failed calls, and unsupported features.

They ask the community whether anyone has built a provider-neutral eval around one workflow, and which variable was hardest to keep constant.

Original post →

More from coding & agent

coding & agent channel →