Guides

Agent Studio

Agent Studio is the visual configuration interface inside the Neuratel dashboard. Every setting available in the UI maps directly to a field in the V3 agent schema, so you can switch between the dashboard and the API at any point without losing work.

Prerequisites: Neuratel account with dashboard access

Opening Agent Studio

Log into the Neuratel dashboard.
Click Agents in the left sidebar.
Select an existing agent or click Create Agent.

The editor opens with seven configuration tabs across the top: Brain, Voice, Transcriber, Conversation, Tools, Transfer, and Analytics. Changes are applied when you click Save on each tab.

Brain

The Brain tab controls which language model powers the agent and how it behaves.

Field	Description	API field
Provider	LLM provider	`brain.provider`
Model	Specific model	`brain.model`
System Prompt	Agent instructions and persona	`brain.instructions`
Temperature	Response variability, 0–1	`brain.temperature`
Max Tokens	Maximum tokens per response	`brain.max_completion_tokens`

Available models

Provider	Models
OpenAI	`gpt-5.4`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`
Groq	`llama-3.1-8b-instant`, `meta-llama/llama-4-scout-17b-16e-instruct`, `openai/gpt-oss-20b`
Neuratel	`phantom` (avg 187 ms, no external dependency)

Tips

Temperature 0.5–0.7 is the right range for support and sales agents. Lower values produce more consistent, predictable responses. Higher values produce more varied language.
System prompts under 1500 tokens keep first-token latency low. Put detailed reference data in a knowledge base instead of the prompt.
Phantom (phantom) eliminates third-party LLM latency and is the fastest option when sub-200 ms response times matter.

Voice

The Voice tab configures which text-to-speech engine and voice the agent uses.

Field	Description	API field
Provider	TTS service	`voice.provider`
Model	TTS model	`voice.model`
Voice	Voice ID or name	`voice.voice_id` / `voice.voice`
Speed	Speaking rate, 0.5–2.0	`voice.speed`
Stability	Consistency vs. expressiveness (ElevenLabs)	`voice.stability`
Similarity Boost	Adherence to original voice (ElevenLabs)	`voice.similarity_boost`

Available providers and models

Provider	Model	Avg latency
ElevenLabs	`eleven_flash_v2_5`	71 ms
Cartesia	`sonic-3`	37 ms
Phantom Speech	`phantom-english-speech-preview`	63 ms

Voice preview

Click the play button next to any voice in the list to hear a sample before committing. Previews use your current Speed and Stability settings so what you hear matches what callers will hear.

Tips

Cartesia sonic-3 has the lowest latency (37 ms avg) and is the best choice when conversational responsiveness is the top priority.
ElevenLabs eleven_flash_v2_5 offers the widest voice selection and the most natural prosody for longer utterances.
Phantom Speech is the best option when you want no dependency on external TTS providers.
Speed 1.0–1.1 sounds natural for most agents. Avoid going above 1.3 on phone calls — it becomes hard to follow on compressed audio.

Transcriber

The Transcriber tab configures how caller speech is converted to text.

Field	Description	API field
Provider	STT service	`transcriber.provider`
Model	STT model	`transcriber.model`
Language	BCP-47 language code	`transcriber.language`
VAD Mode	Voice activity detection sensitivity	`transcriber.vad_mode`

Available providers and models

Provider	Models
Deepgram	`nova-3`, `nova-3-medical`
OpenAI	`gpt-4o-mini-transcribe`
Phantom (Coming Soon)	`phantom-stt-v1`

VAD Mode

Voice activity detection (VAD) determines when a caller has stopped speaking. Three modes are available:

Mode	Behaviour
`default`	Standard sensitivity, works well in quiet environments
`aggressive`	Cuts off sooner — good for fast-paced conversations
`sensitive`	Waits longer — good for slow speakers or noisy lines

Tips

nova-3-medical recognises medical terminology more accurately and is the right choice for healthcare agents.
Set language explicitly (en-US, es-ES, fr-FR) rather than leaving it on auto-detect. Auto-detect adds a small latency penalty and can misclassify accented speech.
If callers are being cut off mid-sentence, switch VAD to sensitive.

Conversation

The Conversation tab controls timing, turn-taking, and what the agent says at the start and end of calls.

Field	Description	API field
First Message	What the agent says when the call connects	`conversation.first_message`
Silence Timeout	Seconds of silence before the agent ends the call	`conversation.silence_timeout_seconds`
Max Duration	Hard cap on call length in seconds	`conversation.max_duration_seconds`
Allow Interruptions	Whether callers can interrupt agent speech	`conversation.allow_interruptions`
Interruption Threshold	Sensitivity of interruption detection, 0–1	`conversation.interruption_threshold`
End Call Message	Agent's closing statement before hanging up	`conversation.end_call_message`
Background Audio	Ambient sound played during agent processing	`conversation.background_audio`

First message tips

Keep the first message under 20 words and end with an implicit or explicit invitation to speak:

Code
"Thanks for calling Acme Corp. How can I help you today?"

Avoid lengthy introductions — callers often start talking before the agent finishes, and a long opener delays the actual conversation.

Interruptions

Enable interruptions (allow_interruptions: true) for support and sales agents where a natural back-and-forth is expected. Set the threshold to 0.5 as a starting point and tune based on call recordings — lower values make the agent easier to interrupt, higher values make it more resistant.

Background audio

Optional ambient sound (hold music, office noise) played during the brief processing gap between a caller's turn and the agent's response. This reduces the perception of silence and makes the conversation feel more continuous.

Tools

The Tools tab lets you define custom function-calling tools and connect external MCP servers.

Section	Description	API field
Function Tools	Custom tools with URL endpoints	`tools.items`
MCP Integrations	Connected MCP servers	configured via `POST /v1/integrations`

Adding a function tool

Click Add Tool.
Enter the tool name, description, and URL.
Define the parameters as a JSON Schema object.
Click Save.

The agent's LLM reads the tool's description and parameters to decide when and how to call it. Write descriptions in plain language that describes what the tool does and when it should be used.

Connecting an MCP server

Go to Settings → Integrations → MCP Servers.
Click Add MCP Server and paste your server URL.
After connecting, return to the agent's Tools tab and enable the integration.
Browse and enable individual tools from the MCP server's tool list.

View available tools from a connected MCP server:

Code
GET /v1/integrations/mcp/{integration_id}/tools

Tips

Write tool descriptions as action phrases: "Look up the status of a customer order by order ID" is clearer than "Order lookup."
Limit to 10 tools per agent. More tools increase the chance of the LLM selecting the wrong one.
Test each tool individually using the Test Tool button before running a full call test.

Transfer

The Transfer tab configures how the agent routes calls to human agents or other numbers.

Field	Description	API field
Transfer Mode	`blind` or `warm`	`transfer.mode`
Transfer Destinations	Number label + phone number pairs	`transfer.destinations`
Transfer Prompt	Instructions for when/how to transfer	`transfer.prompt`

Blind transfer

The agent announces the transfer and immediately bridges the call to the destination. The human agent receives the call with no prior context from the agent.

Warm transfer

The agent stays on the line briefly, delivers a handoff summary to the human agent, and then drops off. This gives the receiving agent context before they start speaking to the caller.

Defining destinations

Add named destinations so the LLM can select the right one based on the conversation:


Code
{
  "transfer": {
    "mode": "blind",
    "destinations": [
      { "label": "Billing team", "number": "+14155550111" },
      { "label": "Technical support", "number": "+14155550222" },
      { "label": "Sales", "number": "+14155550333" }
    ],
    "prompt": "Transfer to the billing team for payment issues, technical support for product problems, and sales for upgrade inquiries."
  }
}

Analytics

The Analytics tab configures what Neuratel records and analyses after each call.

Field	Description	API field
Enable Recording	Record audio of the call	`analytics.recording_enabled`
Success Evaluation	Boolean criteria for call success	`analytics.success_evaluation`
Structured Output	JSON fields to extract from the transcript	`analytics.structured_output`
Summary	Generate a plain-language call summary	`analytics.summary_enabled`

Success evaluation

Write a natural-language description of what a successful call looks like. Neuratel evaluates the transcript against this description after every call and returns a success: true/false flag.

Code
The call is successful if the customer's question was answered or the caller was transferred to the appropriate team. The call is not successful if the caller hung up in frustration or the agent failed to understand the request.

Structured output

Define fields you want extracted from every transcript. The extraction runs after the call and is available via GET /v1/voice-sessions/{id} in the analysis field.


Code
{
  "analytics": {
    "structured_output": {
      "customer_name": "string",
      "issue_category": ["billing", "technical", "other"],
      "resolution": "string",
      "sentiment": ["positive", "neutral", "negative"]
    }
  }
}

Saving and publishing changes

All changes in Agent Studio are saved per-tab when you click Save. Each save increments the agent's version number. Changes take effect immediately for new calls — calls already in progress are not affected.

To revert to an earlier version, use the version history panel accessible from the agent's Settings menu or via the API: POST /v1/agents/{id}/versions/{version}/restore.

Last modified on November 7, 2026

Agent Lifecycle Conversation Behavior

Guides

Agent Studio

Prerequisites: Neuratel account with dashboard access

Opening Agent Studio

Log into the Neuratel dashboard.
Click Agents in the left sidebar.
Select an existing agent or click Create Agent.

Brain

The Brain tab controls which language model powers the agent and how it behaves.

Field	Description	API field
Provider	LLM provider	`brain.provider`
Model	Specific model	`brain.model`
System Prompt	Agent instructions and persona	`brain.instructions`
Temperature	Response variability, 0–1	`brain.temperature`
Max Tokens	Maximum tokens per response	`brain.max_completion_tokens`

Available models

Provider	Models
OpenAI	`gpt-5.4`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`
Groq	`llama-3.1-8b-instant`, `meta-llama/llama-4-scout-17b-16e-instruct`, `openai/gpt-oss-20b`
Neuratel	`phantom` (avg 187 ms, no external dependency)

Tips

Temperature 0.5–0.7 is the right range for support and sales agents. Lower values produce more consistent, predictable responses. Higher values produce more varied language.
System prompts under 1500 tokens keep first-token latency low. Put detailed reference data in a knowledge base instead of the prompt.
Phantom (phantom) eliminates third-party LLM latency and is the fastest option when sub-200 ms response times matter.

Voice

The Voice tab configures which text-to-speech engine and voice the agent uses.

Field	Description	API field
Provider	TTS service	`voice.provider`
Model	TTS model	`voice.model`
Voice	Voice ID or name	`voice.voice_id` / `voice.voice`
Speed	Speaking rate, 0.5–2.0	`voice.speed`
Stability	Consistency vs. expressiveness (ElevenLabs)	`voice.stability`
Similarity Boost	Adherence to original voice (ElevenLabs)	`voice.similarity_boost`

Available providers and models

Provider	Model	Avg latency
ElevenLabs	`eleven_flash_v2_5`	71 ms
Cartesia	`sonic-3`	37 ms
Phantom Speech	`phantom-english-speech-preview`	63 ms

Voice preview

Click the play button next to any voice in the list to hear a sample before committing. Previews use your current Speed and Stability settings so what you hear matches what callers will hear.

Tips

Cartesia sonic-3 has the lowest latency (37 ms avg) and is the best choice when conversational responsiveness is the top priority.
ElevenLabs eleven_flash_v2_5 offers the widest voice selection and the most natural prosody for longer utterances.
Phantom Speech is the best option when you want no dependency on external TTS providers.
Speed 1.0–1.1 sounds natural for most agents. Avoid going above 1.3 on phone calls — it becomes hard to follow on compressed audio.

Transcriber

The Transcriber tab configures how caller speech is converted to text.

Field	Description	API field
Provider	STT service	`transcriber.provider`
Model	STT model	`transcriber.model`
Language	BCP-47 language code	`transcriber.language`
VAD Mode	Voice activity detection sensitivity	`transcriber.vad_mode`

Available providers and models

Provider	Models
Deepgram	`nova-3`, `nova-3-medical`
OpenAI	`gpt-4o-mini-transcribe`
Phantom (Coming Soon)	`phantom-stt-v1`

VAD Mode

Voice activity detection (VAD) determines when a caller has stopped speaking. Three modes are available:

Mode	Behaviour
`default`	Standard sensitivity, works well in quiet environments
`aggressive`	Cuts off sooner — good for fast-paced conversations
`sensitive`	Waits longer — good for slow speakers or noisy lines

Tips

nova-3-medical recognises medical terminology more accurately and is the right choice for healthcare agents.
Set language explicitly (en-US, es-ES, fr-FR) rather than leaving it on auto-detect. Auto-detect adds a small latency penalty and can misclassify accented speech.
If callers are being cut off mid-sentence, switch VAD to sensitive.

Conversation

The Conversation tab controls timing, turn-taking, and what the agent says at the start and end of calls.

Field	Description	API field
First Message	What the agent says when the call connects	`conversation.first_message`
Silence Timeout	Seconds of silence before the agent ends the call	`conversation.silence_timeout_seconds`
Max Duration	Hard cap on call length in seconds	`conversation.max_duration_seconds`
Allow Interruptions	Whether callers can interrupt agent speech	`conversation.allow_interruptions`
Interruption Threshold	Sensitivity of interruption detection, 0–1	`conversation.interruption_threshold`
End Call Message	Agent's closing statement before hanging up	`conversation.end_call_message`
Background Audio	Ambient sound played during agent processing	`conversation.background_audio`

First message tips

Keep the first message under 20 words and end with an implicit or explicit invitation to speak:

Code
"Thanks for calling Acme Corp. How can I help you today?"

Avoid lengthy introductions — callers often start talking before the agent finishes, and a long opener delays the actual conversation.

Interruptions

Background audio

Tools

The Tools tab lets you define custom function-calling tools and connect external MCP servers.

Section	Description	API field
Function Tools	Custom tools with URL endpoints	`tools.items`
MCP Integrations	Connected MCP servers	configured via `POST /v1/integrations`

Adding a function tool

Click Add Tool.
Enter the tool name, description, and URL.
Define the parameters as a JSON Schema object.
Click Save.

The agent's LLM reads the tool's description and parameters to decide when and how to call it. Write descriptions in plain language that describes what the tool does and when it should be used.

Connecting an MCP server

Go to Settings → Integrations → MCP Servers.
Click Add MCP Server and paste your server URL.
After connecting, return to the agent's Tools tab and enable the integration.
Browse and enable individual tools from the MCP server's tool list.

View available tools from a connected MCP server:

Code
GET /v1/integrations/mcp/{integration_id}/tools

Tips

Write tool descriptions as action phrases: "Look up the status of a customer order by order ID" is clearer than "Order lookup."
Limit to 10 tools per agent. More tools increase the chance of the LLM selecting the wrong one.
Test each tool individually using the Test Tool button before running a full call test.

Transfer

The Transfer tab configures how the agent routes calls to human agents or other numbers.

Field	Description	API field
Transfer Mode	`blind` or `warm`	`transfer.mode`
Transfer Destinations	Number label + phone number pairs	`transfer.destinations`
Transfer Prompt	Instructions for when/how to transfer	`transfer.prompt`

Blind transfer

The agent announces the transfer and immediately bridges the call to the destination. The human agent receives the call with no prior context from the agent.

Warm transfer

The agent stays on the line briefly, delivers a handoff summary to the human agent, and then drops off. This gives the receiving agent context before they start speaking to the caller.

Defining destinations

Add named destinations so the LLM can select the right one based on the conversation:


Code
{
  "transfer": {
    "mode": "blind",
    "destinations": [
      { "label": "Billing team", "number": "+14155550111" },
      { "label": "Technical support", "number": "+14155550222" },
      { "label": "Sales", "number": "+14155550333" }
    ],
    "prompt": "Transfer to the billing team for payment issues, technical support for product problems, and sales for upgrade inquiries."
  }
}

Analytics

The Analytics tab configures what Neuratel records and analyses after each call.

Field	Description	API field
Enable Recording	Record audio of the call	`analytics.recording_enabled`
Success Evaluation	Boolean criteria for call success	`analytics.success_evaluation`
Structured Output	JSON fields to extract from the transcript	`analytics.structured_output`
Summary	Generate a plain-language call summary	`analytics.summary_enabled`

Success evaluation

Write a natural-language description of what a successful call looks like. Neuratel evaluates the transcript against this description after every call and returns a success: true/false flag.

Code
The call is successful if the customer's question was answered or the caller was transferred to the appropriate team. The call is not successful if the caller hung up in frustration or the agent failed to understand the request.

Structured output

Define fields you want extracted from every transcript. The extraction runs after the call and is available via GET /v1/voice-sessions/{id} in the analysis field.


Code
{
  "analytics": {
    "structured_output": {
      "customer_name": "string",
      "issue_category": ["billing", "technical", "other"],
      "resolution": "string",
      "sentiment": ["positive", "neutral", "negative"]
    }
  }
}

Saving and publishing changes

To revert to an earlier version, use the version history panel accessible from the agent's Settings menu or via the API: POST /v1/agents/{id}/versions/{version}/restore.

Last modified on November 7, 2026

Agent Lifecycle Conversation Behavior

Opening Agent Studio

Brain

Available models

Tips

Voice

Available providers and models

Voice preview

Tips

Transcriber

Available providers and models

VAD Mode

Tips

Conversation

First message tips

Interruptions

Background audio

Tools

Adding a function tool

Connecting an MCP server

Tips

Transfer

Blind transfer

Warm transfer

Defining destinations

Analytics

Success evaluation

Structured output

Saving and publishing changes

Related

Opening Agent Studio

Brain

Available models

Tips

Voice

Available providers and models

Voice preview

Tips

Transcriber

Available providers and models

VAD Mode

Tips

Conversation

First message tips

Interruptions

Background audio

Tools

Adding a function tool

Connecting an MCP server

Tips

Transfer

Blind transfer

Warm transfer

Defining destinations

Analytics

Success evaluation

Structured output

Saving and publishing changes

Related