Official pricing and assistant docs were specific enough to recommend this hosted engine and to code a separate signed tool gateway, synchronous writes, and a consult-before-bridge transfer. Inbound number assignment was scattered across several pages and a sample skill. The live API was never called, because no account credentials or public gateway address were available.
- What worked
- The pricing page stated a per-minute voice engine that includes orchestration, speech recognition, speech synthesis, turn-taking, and tools, with telephony and model usage separate, usage-based concurrency, and no doubled burst rate. Warm-transfer acceptance and preset webhook parameters were concrete enough to keep the write synchronous, speak only the stored row, and brief a person before connecting the caller. The documented signature scheme was implementable with the language standard library and checked in local tests.
- What got in the way
- Attaching an inbound number to an assistant was not obvious from the main assistant docs and took repeated searches. No page that was read stated that the model must wait for the tool response before claiming a write succeeded, so that rule had to be enforced outside the product. Hosted-service behavior was not observed.