Get started
Concepts
The objects and terms the API uses, and how they relate.
Workspace, members and keys
A workspace (an account in the API) owns everything: agents, numbers, calls, the wallet. People log in to it as members with a role: viewer, member, admin or owner. Programs use API keys, which act with the admin role. A workspace can create sub-accounts that have their own keys and data but spend the parent's wallet. See Team and security and Sub-accounts.
Agent
An agent is the definition of a voice assistant: its welcome message, its prompts, and one or more tasks. The conversation task carries the settings for the three model stages and the call:
| Block | Holds |
|---|---|
transcriber | Speech-to-text: provider, model, language, keywords. |
llm_agent | The language model, and whether the agent is a simple prompt or a graph. |
synthesizer | Text-to-speech: provider, voice. |
api_tools | Function tools the model may call. |
task_config | Call behaviour: silence hang-up, maximum length, voicemail, keypad, and more. |
Every save creates a version; an execution remembers the version it ran on. See Agents.
Graph agent
A graph agent is an agent whose conversation is a set of nodes (steps) joined by edges (conditions). The engine moves from node to node as the caller answers, and reports the path it took. Use it when the order of steps matters. See Graph agents.
Execution
An execution is one call: the record of what was dialled, what was said and what it cost. It moves through statuses such as queued, scheduled, ringing, in-progress and ends in one of completed, no-answer, busy, failed, stopped, balance-low or error. The exact list, and which ones are final, is in Executions.
Engine
The engine is what holds the conversation.
| Engine | What it is |
|---|---|
mock | A scripted conversation with no audio and no dialling. For development and tests. |
own | The platform's real-time voice engine. Needs a telephony provider (or the browser voice test) and model servers. |
The engine is chosen per call (agent_data.engine), else per agent (agent_config.engine), else by the deployment's default. See Voice engine.
Telephony mode
Independently of the engine, the deployment runs telephony in one of two modes. In mock mode, searching, buying and dialling numbers are simulated. In live mode they go to your provider. Development helpers such as the email outbox are switched off in live mode.
Batches, workflows and campaigns
A batch is a CSV of contacts called by one agent. A workflow is a sequence of steps (calls, waits, messages, branches) for one contact. A workflow campaign runs a workflow for a list of contacts. See Batches and Campaigns and workflows.
Post-call data
After a call, the worker produces the transcript, an optional summary, the extracted data you defined with dispositions or extraction categories, the cost breakdown, and the wallet charge. See Post-call data.
Wallet and price book
Calls are paid from a prepaid wallet. The cost of a call is its minutes multiplied by per-minute rates from the price book, a configuration file of the deployment. See Billing.
Compliance settings
Calling hours, a do-not-call list, consent, a daily cap per number and an AI disclosure line are checked by the platform before a call is dialled. They are tools; the legal responsibility is yours. See Compliance toolkit.