Gemini MCP integration

Because hard-coded model IDs go stale. List Models returns the Gemini and Veo models currently available with their capabilities and limits, and preview Veo models in particular can be unavailable or return no video URI, so Ayro verifies availability and prefers a stable model over a preview name someone wrote down last quarter.

Typically 30 to 180 seconds, and occasionally up to around twelve minutes. Ayro waits before its first poll and then polls at intervals rather than treating a slow job as a failure, which is why a video request behaves as a background task rather than an instant reply.

It can size it. Count Tokens applies Gemini tokenization to the input and returns the token count, which is what you need for estimating cost and for checking a long input against a model's limit before the generation call is made.

An operation marked done with no video file means the safety filter rejected it, and the reason is reported in the response rather than guessed at, so the prompt is revised and a fresh job started - a failed operation name is never reused. Heavy concurrency returns 429 RESOURCE_EXHAUSTED, so Ayro backs off instead of retrying immediately.