Skip to main content

LLM and model settings

An agent's behavior depends on two sets of model settings:

  • The Models card in the agent's Advanced mode Settings panel, where you choose which model connections the agent uses.
  • The connection itself, created in Settings > Connectivity, where you enter the provider URL, key, model name, and sampling settings.

This page explains each setting — what it controls, when to change it, and a safe starting point. For the steps to select a model on an agent, see Select AI models; for the steps to create a connection, see Add an LLM integration.

The Models card

The Models card lists one row per model role. Each row is a dropdown of the matching connections from Settings > Connectivity; the placeholder reads None until you pick one. Selecting a connection here attaches it to this agent — it does not create or edit the connection.

RowWhat it doesWhen you set it
LLMThe language model that understands customer messages and generates replies.Always — the agent does not run without it.
MCP ToolsTools provided by Model Context Protocol servers. You can select more than one connection.When the agent calls tools from an MCP server.
OpenAPI ToolsHTTP tools generated from an OpenAPI specification. You can select more than one connection.When the agent calls tools created from an OpenAPI spec.
Visual Language ModelA model that combines visual understanding with text processing.When the agent must interpret images.
Text-To-SpeechConverts the agent's replies to spoken audio. You can select more than one connection.Voice agents.
Speech-To-TextConverts the caller's speech to text.Voice agents.
Skills Embedder ModelAn embedding model that turns text into vectors.When the agent's skills need embeddings.
Transcription modelConverts audio recordings into text.Call Analytics agents.

The rows shown depend on the agent type. A Call Analytics agent shows the Transcription model row in place of the live speech rows; the Visual Language Model row appears only for agent types that support images. After you change a selection, click Save, then restart the agent — Advanced mode does not restart it on save.

note

LLM is the only required row. Without a selected LLM the agent cannot start. The other rows are optional and depend on what the agent does.

LLM connection settings

The LLM connection is created under AI Models > Large Language Model (LLM) in Settings > Connectivity. It connects any model behind an OpenAI-compatible API. These are the fields you set on the connection.

FieldRequiredWhat it doesWhen to change it
NameYesA label for the connection, unique within your tenant.Set once. Use a name that identifies the provider and model, for example OpenAI gpt-4o prod.
API URLYesThe OpenAI-compatible base URL of the provider.When you switch providers or endpoints. Use the base URL — for OpenAI, https://api.openai.com/v1.
Access TokenYesThe API key issued by the provider. Stored as a secret and shown masked after saving.When the key is rotated or revoked. Type the new value over the masked field to replace it.
Model NameYesThe model identifier exactly as the provider expects it, for example gpt-4o.When you move to a different model, or your provider deprecates the current one.
TemperatureYesHow varied the model's replies are. See Temperature.Lower for fact-based agents; raise only if replies feel too rigid.
JSON paramsNoAdditional sampling parameters as a JSON object, sent to the provider with each request, for example {"max_tokens": 1024}. Defaults to {}.Only when your provider documents a parameter you need. Leave empty if unsure.

After you save the connection, the portal tests it by sending a real request to the model, and the connector card shows a green status dot when the test succeeds. See Check integration status.

Temperature

Temperature controls how varied the model's output is: low values keep replies focused and repeatable, high values make them more varied and less predictable.

ValueModel behavior
0.0Accurate and deterministic. Best for support agents and anything fact-based.
0.7Balanced between accuracy and variety — works for most conversational agents.
1.0Highly varied and less predictable. Suited to brainstorming and marketing copy.

Start with a low temperature (0 to 0.3) for customer-support agents and raise it only if replies feel too rigid. Use the range your model provider documents.

Speech model settings for voice agents

A voice agent needs two speech connections in addition to its LLM: one to transcribe the caller and one to read replies aloud. Create them in Settings > Connectivity, then select them in the Speech-To-Text and Text-To-Speech rows of the Models card.

Speech-to-text (STT) — created under AI Models > Speech-to-text (STT).

FieldRequiredWhat it does
API URLYesThe OpenAI-compatible base URL of the speech-to-text service.
Access tokenNoThe API key issued by the provider. Stored as a secret.
TypeNoThe transcription engine, for example whisper or local_whisper.
Model nameNoThe transcription model identifier the provider expects.
CertificateNoA client certificate, when the service requires one. Stored as a secret.

Text-to-speech (TTS) — created under AI Models > Text-to-speech (TTS).

FieldRequiredWhat it does
API URLYesThe OpenAI-compatible base URL of the text-to-speech service.
Access tokenNoThe API key issued by the provider. Stored as a secret.
TypeYesThe synthesis engine, for example open_voice, azure_speech, or eleven_labs.
RegionNoThe provider region, when the service uses one.
VoiceNoThe voice the agent speaks in.
Speech speedNoThe speaking rate.

Speech-to-text connections are health-checked and show a status dot; text-to-speech connections are not, so confirm one works by placing a test call. The SIP voice channel uses these speech connections; the Twilio channel uses Twilio's built-in speech and does not need them. For the full field reference and setup steps, see Add speech integrations for voice agents.

Was this article helpful?