Agent
A configured AI participant that joins a phone call. An agent has a voice, an LLM, a system prompt (purpose, personality, instructions, guardrails), zero or more knowledge bases attached for retrieval, and zero or more tools it can invoke during the call. Agents are versioned. You edit a draft, then publish to make it live on the assigned phone numbers.Conversation
A single interaction between an agent and one or more participants — a phone call, a web chat or voice session, or an email thread. Each conversation produces a transcript and metadata (duration, outcome, caller ID), plus an audio recording for voice. See Conversations.Channel
The medium a conversation happens over. Anyreach supports four:telephone (phone calls), webrtc (web voice and chat), email, and text. One agent can serve several channels, with per-channel overrides. See Channels.
Knowledge base
A collection of sources (uploaded files or crawled URLs) that have been chunked and embedded into a vector index. Agents query KBs via retrieval — given a user question, the top matching chunks are pulled into the LLM context for that turn. A KB is identified by adataset_id. Internally KBs are sometimes called datasets; in the UI you’ll always see “Knowledge Base.”
Workflow
A directed graph of steps that processes input and produces output. Workflows are general-purpose automation: they can be invoked by an agent as a tool, by a webhook, by a schedule, or directly via API. Workflows are versioned. Only published versions can be referenced from agents or external triggers.Step
A single node in a workflow. Anyreach ships eight step types —input, code, condition, http, app, output, wait, note — each with its own configuration and behavior. See the step reference.
Trigger
The event that starts a workflow execution. Triggers come in four flavors: manual (run from the UI or API), scheduled (cron), webhook (external service via Pipedream), and agent-embedded (an agent calls the workflow as a tool, or fires it pre/post-call).Expression
Inline templating used inside step configuration. Two flavors:{{...}}— JSONata, for deterministic data transformations${...}— AI expression, where the contents are sent to an LLM and replaced with the result

