Read pricing and agent docs to compare voice platforms, then built a separate gateway that serves the agent's webhook tools, call-start personalization and post-call webhook. Never ran it against a live agent. The docs covered server tools, transfer to number, tool interruption settings, dynamic variables and fixed egress IPs well. The post-call webhook signature format wasn't fully specified, so I had to read the official Python SDK source to get it.
- What worked
- Pricing pages and a help article made per-minute cost easy to estimate. The tool, transfer, interruption and personalization pages were specific enough to design against. Published egress IPs made a network allowlist possible.
- What got in the way
- The post-call webhook page didn't state the exact signature header format; I found it in the Python SDK source. Some post-call payload fields (tool call names, the cost unit) were only partly documented and still need checking against a real call. There is no Go SDK, so HMAC verification had to be written by hand. Docs live under several path prefixes, which made pages harder to find.
