Add vapi-build: build a Vapi agent from an organization's own material

[vapi-eisen]

What is this project?

vapi-build is an Agent Skill plus a small Python host that builds a complete Vapi voice agent from an organization's own material (website, knowledge articles, call transcripts or speech IVR logs, OpenAPI spec): an evidence-linked ontology, a reviewed plan, then the knowledge base, tools, assistants, squad, structured outputs, and simulation suite, with two yes/no gates on one review page. It installs into Claude Code, Codex, and Claude, and the README gives the steps for each.

Checklist

• I've tested this from a fresh clone (pytest, ruff, the VapiAI/skills validator, and the CLI's doctor command all run from projects/vapi-build in this clone)
• I've included no sensitive information including keys, tokens, data
• Project level README.md includes the setup/run instructions
• I've added a SKILL.md if this project can be a repeatable pattern
• I've included an .env.example
• Added a row for this project to the table in the root README.md

Anything reviewers should know?

• Python 3.11+ with jsonschema is the only hard dependency; tests touch no network and use a fake Vapi transport.
• The skill never asks for a key in chat. The CLI copies VAPI_API_KEY from the environment, a shell profile, or a .env file into ~/.config/vapi-build/env without printing it.
• apply creates real Vapi resources and simulate spends credits; both require an explicit yes and there is a teardown that removes everything the build created.
• The folder is a valid skill directory under the VapiAI/skills validator (copied verbatim into scripts/), so it can later move there as an experimental reference workflow. agents/openai.yaml carries the Codex interface metadata.
• Claude Code has had the most exercise. Codex and claude.ai follow their published skill conventions; the README lists what differs (no browser or persistence in claude.ai, network approval in Codex).

🤖 Generated with Claude Code