Start here
What's new, November 2026
A public REST API with TypeScript and Python SDKs, the vai CLI, agent tests in CI, a leaner MCP server, undo, and webhook delivery logs.
Build from your own code
- A public REST API. The operations behind the app, the MCP server and
Builder are now routes under
https://app.verticalai.com.au/api/v1, used with a personal access token: agents, drafts and publishing, tools, variables, knowledge, tests, calls, phone numbers, campaigns, contacts and webhooks. Errors are problem details with a stablecode, aPOSTcan carry anIdempotency-Keyso a retry never runs twice, and lists page with a cursor. See REST API and the API reference. - TypeScript and Python SDKs.
@verticalai.com.au/sdkandvertical-aigive every route a typed function. See the TypeScript SDK and the Python SDK. - The
vaiCLI. Sign in from the browser, keep an agent's config in a repository (vai pull,vai diff,vai push), rerun tests on every save withvai dev, and read calls from the terminal. See The vai CLI. - Agent tests in CI. The
verticalau/vai-agent-testsGitHub Action runs an agent's tests on every pull request and fails the job when too few pass, with a spend limit. See Agent tests in CI.
Safer changes
- Undo. Before each change made through Builder, your AI assistant, the
API or
vai push, the draft is saved. Say "undo that" to Builder, runvai undo, or callundo_last_changeto put the draft back, up to ten changes. See Undo. - Fix with Builder. An ended call, a failed test run and a readiness issue each have a Fix with Builder button that opens Builder on the problem. See Fixing a call.
- Approvals you can trust. A change that reaches live calls, deletes something or spends money asks first, everywhere: in Builder, in your AI assistant, over the API and in the CLI. The audit log records who approved each one.
Connecting an AI assistant
- Read-only by default. A new connection starts read-only. Editing, deleting and managing the workspace are separate choices on the consent screen. See Connect your assistant.
- A smaller tool list. The MCP server lists the tools for building, testing and shipping, and finds the rest by search, so it fits in any assistant's context.
- Guides and docs as resources. Assistants that show MCP resources can read the prompt-writing guide, the build guide, every docs page and your agents' recent calls.
Tokens and webhooks
- Token expiry. A personal access token now lasts 30 days, 90 days or a year, and you choose when you make it under Settings > API tokens.
- Webhook delivery log. Every event sent to a webhook is logged with each attempt, and a missed event can be sent again once your server is fixed. See The delivery log.
Related
Getting started
You sign up and take a real call in about thirty minutes. The sign-up wizard builds your first agent and lets you talk to it in your browser. Test calls are free.
Core concepts
An agent is one language model, one system prompt and a flat list of tools. The workspace owns the phone numbers, and the published version answers calls.