Configuration fields
System prompt
The system prompt defines your agent’s persona, goals, tone, and guardrails. For voice agents it’s the primary behavioral lever — write it for speech, not screen. See the System prompt guide for detailed authoring guidance.Prompt timezone
Set a prompt timezone to make the agent aware of the current date and time. When set, Zeptar appends a single line — the current date and time for that zone — to the system prompt at the start of every conversation, so the agent can answer questions like “what’s today’s date?” or apply business-hours logic.- Stored as a zone, evaluated at dispatch. The setting holds an IANA timezone (for example
Europe/Paris), never a frozen timestamp. The actual date and time are computed when each conversation begins, so the agent always sees the real “now.” - Off by default. Leave it unset and nothing is added to the prompt.
- Per-conversation override. Pass
prompt_timezonein the conversation overrides to use a different zone for a single conversation, ornullto turn the date/time line off for that conversation.
First message
The first message is the opening line the agent speaks when a conversation starts. It frames the call and sets the caller’s expectations immediately.- Filled in — the agent speaks this line before the caller says anything.
- Left empty — the agent waits for the caller to speak first. After the wait time elapses with no caller speech, the agent opens the conversation itself (generated from the system prompt).
First message per language
When the agent has additional languages, a language menu appears in the first message footer. Pick a language to see and edit the line the agent opens with when a call runs in that language. The first time you open a language, its message is translated from the default first message automatically, and you can then edit it. If you change the default first message or the default language later, opening a language translates it again, unless you’ve edited its message yourself, in which case your text is kept. A language without a message opens with the default first message. Translate to all translates the default first message into every additional language in one step, replacing any per-language messages already written.{{variables}} are kept as they are. Review the translations before publishing; you can edit any of them by hand.
In the API, per-language messages live in language_first_messages, keyed by additional-language code ({"de": "Hallo! Wie kann ich helfen?"}). language_first_message_translations records the default message and language each translation came from, which is how the editor knows a translation is out of date. Removing a language from the agent also removes its message.
Wait time
When the first message is empty, wait time controls how many seconds the agent waits for the caller to speak before it opens the conversation on its own.- A number — wait that many seconds, then speak.
0opens immediately. -1— wait indefinitely; the agent never speaks first and always lets the caller lead.- Left blank — use the default of 7 seconds.
Interruptible
When Interruptible is on (the default), the caller can talk over the agent and the agent stops to listen. Turn it off for flows where the agent must finish speaking — for example reading a required disclosure — before it accepts input.Language
The language setting controls which language the agent uses for speech recognition and synthesis.- Default language — the primary language for the conversation. The agent speaks in this language and optimises speech recognition for it.
- Additional languages — optional list of secondary languages. When provided, the agent can detect and switch to any of the listed languages mid-conversation.
Setting a narrower language improves transcription accuracy for callers who
speak that language. If your callers always speak a single language, set only
the default and leave additional languages empty.
Transcriber
The transcriber is the speech-to-text model that turns what the caller says into text for the LLM. It’s saved asstt with a provider, a model and optional settings.
The catalog (GET /v1/agent-config/catalog, under stt) lists every transcriber with its price per minute, the languages it handles (multilingual, single or english) and the settings it takes. Which settings are available depends on the model:
Saving a setting the model doesn’t take is rejected with
unknown_stt_setting; an unknown model with unknown_stt_model.
The transcriber has to handle the agent’s languages. With additional
languages, pick a
multilingual model; English-only models fit only an
English default language. If the saved transcriber doesn’t fit, calls use
Deepgram Nova-3 instead, and the editor flags the mismatch.LLM
The LLM setting selects the model that drives the agent’s reasoning and response generation.
Zeptar’s model catalog spans multiple providers. You can pick any model listed in
GET /v1/agent-config/catalog without changing anything else in your setup — Zeptar handles routing to the selected provider.
Leaving temperature blank tells Zeptar not to send a temperature parameter at
all. The model runs with its own internal default, which may differ from what
you’d get by setting 1.0 explicitly. If you need deterministic or creative
output, set an explicit value.
Voice
The voice setting selects the synthesis voice the agent speaks with.
Like models, the voice catalog spans multiple providers. Browse available voices in
GET /v1/agent-config/catalog and switch at any time — the change takes effect on the next published version.
How to update your agent’s configuration
1
Open your agent
Navigate to Agents in the sidebar and select the agent you want to configure.
2
Go to the Agent tab
Inside the agent detail page, select the Agent tab. This is where system
prompt, first message, language, model, and voice all live.
3
Make your changes
Edit any field. Changes autosave to your draft on the current branch as you
type — there’s no explicit save button.
4
Publish to create a version
When your changes are ready, click Publish. Publishing commits your draft as a new immutable version on the current branch. Conversations that start after this point run against the new version.
Configuration and versioning
Every configuration field — system prompt, first message, language, LLM, voice — is part of the versioned agent config. This means:- Edits autosave to your draft. Nothing reaches production until you publish.
- Publishing creates an immutable version. Once created, a version is never changed or deleted.
- Conversations pin to a version. A conversation started at version 3 stays on version 3 for its entire lifetime, even after you publish version 4.
- Branches let you experiment safely. Fork a branch from any version, make bold changes to the configuration, publish, and test —
Mainis untouched until you merge.