Assistants

Each directory directly under the repository root that contains config.yaml defines an assistant. Its directory name identifies it across syncs.

.
├── reception/
│   ├── config.yaml
│   ├── base.md
│   ├── default.md
│   ├── functions.yaml          # optional
│   └── workflows/             # optional
└── sales/
    ├── config.yaml
    ├── base.md
    └── default.md

Assistant directory names must start with a letter or digit, contain only letters, digits, ., _, or -, and be at most 100 characters long. Put assistants directly at the root: assistants/reception/config.yaml is not discovered as an assistant. A repository-level config.yaml is also ignored.

Configuration

name: Reception
type: inbound
spokenName: Reception
timeout:
  firstNotice: 30
  final: 60
languages: [en, nl]
timezone: Europe/Amsterdam
opening:
  en: "Hello, you are speaking with {{assistant_name}} at {{organization_name}}. How can I help?"
  nl: "Hallo, u spreekt met {{assistant_name}} van {{organization_name}}. Waarmee kan ik u helpen?"
faq:
  hours:
    en:
      question: "When are you open?"
      answer: "Monday to Friday, from 9am to 5pm."
    nl:
      question: "Wanneer zijn jullie open?"
      answer: "Van maandag tot vrijdag, van 9 tot 17 uur."

Unknown fields are rejected. Keep prompts in Markdown files and functions in functions.yaml. Phone assignments, transfer settings, SMS sender settings, and integration credentials are configured in the dashboard.

Models and voices

Set models to choose the same approved options shown in the dashboard's Models & voice page. Changes apply to new calls. Speech recognition stays on Deepgram Nova-3.

name: Reception
timeout: { firstNotice: 30, final: 60 }
languages: [en, nl]
timezone: Europe/Amsterdam
models:
  llm: openai/gpt-4.1-mini
  voices:
    en:
      tts: gradium/default
      voice: 4SZHfMpw-p46Ywgs
    nl:
      tts: cartesia/sonic-3.6
      voice: 9626c31c-bec5-4cca-baa8-f8ba9e84c8bc

When models is provided, llm and both voices language entries are required. Each language has its own tts model and voice ID, with an optional name for display. Providers only need to support the language they are selected for. Existing configurations with a shared models.tts, string voice IDs, and optional voiceNames remain supported.

Omit models to initialize a new assistant with the initial AI model and concrete speech voices. Later syncs preserve that assistant's saved selections when models is omitted. Changing the platform's initial voices does not change existing assistants. Every voice must be a concrete provider ID; default is not a voice ID. The dashboard library searches current provider catalogs by name, ID, or descriptive keywords. Copy IDs from that library and preview them before using them in Git. Sync checks model names, voice ID syntax, and model language support; provider availability is checked when speech is generated. An ElevenLabs shared voice must already be accessible to the platform account. Gradium can handle English while another provider handles Dutch, as shown above.

Supported timezones

Opening messages and FAQ

Every language used in opening or faq must also appear in languages. Opening messages support {{organization_name}} and {{assistant_name}}; see Prompt variables.

The example above uses a stable key, hours, with a translated question and answer for each language. If the question is shared across languages, a FAQ entry can instead use the question itself as the key:

faq:
  "Wi-Fi?":
    en: "Ask reception for the guest network password."
    nl: "Vraag de receptie naar het wachtwoord van het gastnetwerk."

FAQ keys and questions must be nonempty and at most 500 characters. Answers must be nonempty and at most 10,000 characters. FAQ supplies knowledge for the assistant to use during the conversation.

Multiple assistants and identity

Add another root directory with its own config.yaml, base.md, and default.md to create another assistant. Each assistant has its own prompts, functions, workflows, and dashboard integration connections.

Changing name updates the existing assistant's display name. Renaming its directory creates a new assistant and marks the previous one as removed from the repository. The old assistant's phone assignments do not move to the new one.

An existing assistant's type cannot change between inbound and outbound. Create a new directory for the other type. See Syncing and changes for identity, removals, and import limits.