title: Global settings description: Flow-wide settings in the Global settings tab: voice and language, model, temperature and speed tiers, the global prompt, flex or rigid mode, the welcome message, functions, MCP servers and knowledge bases. order: 2#
The Global settings tab in the flow builder's right column holds everything that applies to the whole flow, plus the agent's own settings. Click Global settings at the top of the right column to open it. Selecting a node switches the column to Node settings; closing the node's settings switches back.
Agent settings#
| UI label | Field | Stored on |
|---|---|---|
| Voice & Language | language, voice_id, voice_model |
Agent |
Model#
Global Settings shows the model picker and an LLM options button.
| UI label | Field | Default | Notes |
|---|---|---|---|
| Model picker | model_choice.model |
gpt-5.6-terra |
Cascading model for every node without a node override |
| Speech-to-speech models in the picker | model_choice: {type: "s2s", s2s_model} |
Replaces STT, LLM and TTS with a realtime model Needs keys | |
| LLM Temperature | model_temperature |
0 |
0 to 1 |
| Fast tier | model_choice.high_priority |
off | "1.5x LLM price, lower latency." |
| Ultrafast | model_choice.ultrafast |
off | Fast tier plus the lowest-latency serving for the model. Turning it on also turns on Fast tier |
Fast tier and Ultrafast are not available for Claude models; the options panel says so and switching to a Claude model turns them off.
{ "model_choice": { "type": "cascading", "model": "gpt-5.6-terra", "high_priority": true, "ultrafast": false }, "model_temperature": 0 }Nodes can override the model under Node settings. See Nodes and edges.
Global Prompt#
Global Prompt (global_prompt) "Applies to every node: personality, background, rules." It comes first in the system prompt on every node, before the node's "Current Task". Type {{ to insert a dynamic variable. The editor counts characters and estimated tokens.
Put here what is true for the whole call: who the agent is, the business, tone, rules ("never quote prices"), and facts every node needs. Put step-specific instructions in nodes.
Transition Flexibility#
| Option | flex_mode |
Behaviour |
|---|---|---|
| Flex Mode | true |
"The agent moves between steps on its own, which handles callers who jump around." The flow is compiled into one prompt |
| Rigid Mode | false or unset |
"Transitions only fire on the conditions you set, in order." |
See Flex mode for what changes at runtime.
Welcome Message#
| UI | Field | Notes |
|---|---|---|
| AI speaks first / User speaks first | start_speaker |
Default agent |
| (AI speaks first) | "The opening line is the start node in the flow. Edit it there (Prompt or Static Sentence)." A static sentence on the start node is spoken as written; a prompt is generated | |
| Pause Before Speaking | begin_message_delay_ms on the agent |
0 to 5 s |
| Speak first if the caller stays silent (User speaks first) | begin_after_user_silence_ms |
Off by default. When on, 0.5 to 15 s, starting at 3 s |
Flows have no separate welcome message field: the start node is the welcome.
Settings sections#
Below the welcome message are the same collapsible sections as other agents. The first three belong to the flow:
| Section | Flow fields |
|---|---|
| Functions | tools[]. Each tool gets a tool_id used by function and subagent nodes |
| MCP Servers | mcps[], used by MCP tools and MCP nodes |
| Knowledge Base | knowledge_base_ids, kb_instruction (Retrieval instruction), kb_config.top_k (Chunks to retrieve, 1 to 10, default 3) and kb_config.filter_score (Similarity Threshold, default 0.6) |
The rest are agent settings: Speech Settings, Realtime Transcription Settings, Call Settings, Agent Settings, Post-Call Extraction, Agent Handbook, Workflows, Guardrails, Security & Fallback Settings and Webhook Settings.
Knowledge base retrieval runs on caller turns. A node can use its own knowledge bases and retrieval instruction through Node settings.
Default dynamic variables#
default_dynamic_variables on the flow fill {{variables}} when a call does not provide them. The test panel's { } button edits the values used in test chats and test calls.
Through the API#
All flow-level fields are on the conversation flow object. Update them with PATCH /update-conversation-flow/:id. Voice, language and the other agent sections are on the agent (PATCH /update-agent/:agent_id).