Open the correct workspace before editing. You need permission to view agents and, to change them, permission to edit agents.
1. Find the agent and edit its basics
Open Agents, search by name or ID and select the agent. In Assistant, edit its name or description and check its direction. The Active switch controls whether it can handle calls. The provider cannot be changed after creation.
Search for the agent and open its Assistant settings. 2. Update the greeting or prompt
Open Prompt. Edit First Message for the greeting and System Prompt for the agent's role, goals and boundaries. Review who speaks first, Caller can interrupt the first message and any prompt warnings.
The Prompt tab with the saved greeting and system instructions. 3. Adjust the voice or other settings
Open Voice. On an Inspra Agent, choose Standard or Duplex under Voice Mode. In Standard mode, choose the Voice provider: Inspra voices, Sarvam AI (Indian languages) where offered, or a provider whose key is connected. Select Change Voice to choose another voice and review speech-to-text, language and model settings as needed. Under Advanced on the same card, an Inspra Agent also has Filler words while thinking, which says a short phrase like "one moment" while it thinks.
Voice Mode on the Voice tab: Standard selected, Duplex available, and Callum under Voice Configuration. In Duplex mode, choose Grok Voice or GPT-Live under Realtime model and use Change voice to select a built-in voice. Your Standard settings are kept for switching back. See Duplex models for supported voices, previews and setup.
Use Tools for actions and variables, Knowledge for business information, and Advanced for analysis settings: Call analysis criteria, Data fields and Scorecards. Change only what the agent needs.
4. Save and manage the agent
Click Update Agent at the bottom and confirm the sync badge shows Synced. Use Web Call to test the saved changes; end the call when finished.
The three-dot menu offers Duplicate Agent, Export JSON and Archive. Archive hides an agent from the normal list; enable Show archived to find it again and restore it. Delete permanently is irreversible.
The agent menu, Show archived filter and Voice Mode selector.
Duplex model and voice settings
On an Inspra Agent, Voice Mode offers two ways to listen and respond. Standard configures speech recognition, a language model and speech synthesis separately. Duplex uses one model that listens and speaks, for faster replies, with support for interruptions; the hint under the buttons reads Duplex: one model listens and speaks for faster replies.
| Setting | Standard | Duplex |
|---|---|---|
| Model | Choose an LLM provider and model. | Choose Grok Voice or GPT-Live under Realtime model. |
| Voice | Choose a voice from a compatible synthesis provider. | Choose one of the selected realtime model's built-in voices. |
| Recognition | Choose the Speech-to-Text provider, model and language. | Handled by the realtime model; separate Speech-to-Text controls are hidden. |
Duplex is available for Inspra Agents. Vapi, Retell and ElevenLabs agents use their provider's settings.
Models and voices
The voice picker shows the built-in voices for the selected Realtime model. Search or filter the voices, select a voice to inspect its details, then click Use voice to apply it.
| Realtime model | Default voice | Available voices |
|---|---|---|
| Grok Voice | Ara | Ara, Eve, Leo, Rex, Sal, Atlas, Luna, Iris, Orion, Helix, Carina, Zagan, Altair, Zenith, Perseus, Helios, Lux, Kepler, Rigel (Australian), Cosmo, Celeste, Ursa, Sirius, Lumen, Castor, Naksh |
| GPT-Live | Marin | Marin, Cedar, Beacon, Cinder, Stone, Vesper, Quartz, Ripple |
Changing the realtime model selects that model's default voice: Ara for Grok Voice or Marin for GPT-Live. Choose the voice again if you want a different one. Grok Voice includes Rigel (Australian); GPT-Live includes Quartz and Ripple (Australian).
Every voice has a preview recording: Grok Voice samples come from xAI, and GPT-Live samples were recorded from GPT-Live sessions. Samples can sound different from a live Duplex call, so use Web Call to hear the voice with your agent. Custom-text preview generation is unavailable in the Duplex picker.
After selecting the model and voice, click Update Agent and confirm the agent is Synced. Reopen its voice settings to check the saved selection. Use Web Call with microphone access to test pronunciation, interruptions and any configured tools or transfers. End the call when finished. Live test calls can use credits.
Switch between modes
Your Standard voice, LLM and transcription settings are saved while Duplex is active. To return to them, select Standard, review those settings and click Update Agent. The saved Duplex selection is kept so you can switch back later. Check any required Standard provider connections before testing again.
JSON export and import retain the voice mode and realtime model/voice selection. An imported Duplex agent must include a supported model and a compatible voice.
Troubleshooting
| What you see | What to do |
|---|---|
| Saved Duplex settings are incomplete | Choose a supported realtime model and voice, then save again. Incomplete settings block saving or calling. |
| A model or voice is rejected | Send provider, model and voice together. Check the model's built-in voice set and the lowercase voice ID. |
| No preview sample is available | GPT-Live has no supplied samples. Save the agent and test it with Web Call. |
| The voice changes after choosing another model | Model changes select the new model's default voice. Choose your preferred voice before saving. |
| Voice settings were not saved correctly | Reload the agent, select the model and voice again and save. Verify the saved selection before making a call. |
For programmatic setup, see Duplex API and MCP configuration.







