When Camille writes to her repair shop's WhatsApp number, the shop's automation asks her name, checks whether her bike is ready, and hands her to the team when she asks for a person.
An automation is a graph that messages a contact and reacts to their replies. Each WhatsApp number the workspace connects has one. The shop's manager describes what they want to their agent, the agent writes the graph, and the manager reviews it in the console, beside its number, before anything reaches a customer.
What makes a graph?
Triggers are its entry points:
message.receivedstarts a run when a contact writes to the automation's number, at the step it leads to.Steps send messages, ask questions, use AI, wait, and write values.
Transitions lead from a trigger or a step to the next step. Each reply button or list row has its own.
Every trigger and step has an ID its author chooses, such as ask_name, unique within the graph. A graph has at most 20 triggers and 100 steps, and no transition leads back to an earlier step.
How does an automation reach contacts?
Connecting a number gives it an empty automation, whose draft a member or their agent writes.
They save the draft as a version, with a message saying what changed.
An administrator publishes that version.
A contact who writes to that number starts a run: one contact going through that version.
How does an agent write a draft?
replaceAutomationDraft writes the whole draft at once, in the shape getAutomation returns. Read the draft before replacing it, since a member may have changed it meanwhile, and build it over a few replacements, so a watching member sees it take shape.
Saving refuses a draft that breaks a drafting rule, such as a step no trigger reaches. A saved draft may still hold what publishing would refuse, listed as draftProblems, which the console page shows on their nodes. Its templatePreviews hold what a contact reads of each template a send_template step sends.
How does an automation answer in several languages?
An automation that lists its languages, such as ["en", "fr"], answers each contact in their language when it offers it, and in its first language otherwise. Each text a contact reads is then either one text for every language, such as a link, or one per language, such as { "en": "Hi!", "fr": "Bonjour !" }. A send_template step names its template's language and variables once or per language. The steps, buttons, and transitions stay the same in every language. Publishing refuses a text missing one of the automation's languages.
Heythere does not guess a contact's language. A classify step that tells it from their first message, or a question that asks it, leads to a set_value step that sets contact.language, and the following steps answer in it. The console shows the texts in the language picked in the canvas corner.
In the API
Every operation is a REST endpoint and, unless marked REST only, an MCP tool of the same name. The OpenAPI document describes their parameters, payloads, and errors.
- GET
getAutomationGet an automation
/api/organizations/{organizationId}/workspaces/{workspaceId}/automations/{automationId} - PUT
replaceAutomationDraftReplace an automation's draft
/api/organizations/{organizationId}/workspaces/{workspaceId}/automations/{automationId}/draft - GET
watchAutomationREST onlyWatch an automation
/api/organizations/{organizationId}/workspaces/{workspaceId}/automations/{automationId}/changes