People ask the same question in different places. A parent sends it on WhatsApp. A finance lead types it into your website. Someone else rings the office. If each channel has its own bot with its own copy of the content, the answers drift apart within a month.
In Pragence, a channel is a delivery surface attached to a project. The project holds the sources, the published configuration and the assistant. You configure the assistant once and switch channels on. This post covers the seven channels that answer today and what is shared between them.
What every channel shares
Text conversations on every channel go through the same product query service. It checks that the organization has a grant for the product, uses the project’s current published configuration, and records usage. The assistant then runs the same retrieval over the same sources.
Permissions follow the person asking. Your own staff, signed in to the console, can read documents shared with them. Everyone else, including website visitors, portal users, callers, WhatsApp, email and Slack senders and API keys, gets the documents shared with the whole organization. A file whose sharing was mirrored from Google Drive to three people stays with those three people, whatever the channel.
1. Website chat
The chat widget runs in three layouts: a floating bubble, inline inside an element on your page, or a hosted full-page chat. You create it in the console and copy the generated embed code. The snippet holds a widget key, which identifies the widget. It never contains a secret API key.
Answers in the widget show their citations underneath. Visitors can switch the widget’s language, and the choice is remembered in their browser.
2. Voice in the browser
Visitors can also talk to the assistant from your site. The voice model is Gemini Live.
When a voice session starts, the widget asks the Pragence server for a session. The server builds the voice configuration, including the instructions and the knowledge search tool, and mints a short-lived token with that configuration locked. The browser then connects directly to Gemini with the token. The browser cannot change the instructions or the tools, and the audio never passes through Pragence servers.
Voice uses a read-only knowledge search. Sources and document permissions are resolved live for each session.
Transcripts are off by default. If you switch on transcript saving for the voice channel, the widget sends the text of the conversation when the call ends, and it appears in your inbox alongside text chats. Audio is never stored either way.
3. Phone
Pragence can answer calls on a phone number from your Telnyx or Vonage account. Through the telephony API, an administrator can search the provider’s available numbers, buy one and bind it to a project, list numbers and release them.
On a call, the roles are reversed from browser voice. The provider streams the call audio to Pragence, and a server-side bridge holds the Gemini Live session. The assistant is told the number that was dialled and the caller ID. It keeps its turns to one or two short sentences, never reads out web addresses, and can ask for a name and email address and read them back to confirm. If it cannot help, it offers to take a message with a callback number.
Because the server holds the session, tools run on the server. Actions that need a person’s approval are unavailable on calls and are never confirmed automatically.
With the voice CRM module enabled and its background worker running, ended calls are processed into structured CRM records and follow-ups.
4. WhatsApp
The WhatsApp channel connects to your WhatsApp Business account through Meta’s Cloud API. Each incoming webhook is checked against its signature. The receiving number identifies the project, the project’s assistant answers each text message, and the reply goes back through the Cloud API.
Each sender has a stable thread per channel, so follow-up questions keep their context.
5. Email
Each assistant can have an email inbox. Inbound email arrives through a signed webhook from AgentMail, our email provider. Pragence checks the signature, runs the assistant over the email body and replies in the same thread.
On WhatsApp and email, the reply is the answer text. The numbered markers stay in the text, but the source cards you see in the widget are not attached to the message.
6. Slack
The Slack channel uses your own Slack app. Someone mentions the app in a channel, or sends it a direct message, and the answer arrives in a thread under the question. The sources are listed under the answer, linked when the source has a web address. Replies in the same thread continue the same conversation.
Each request from Slack must carry a fresh signature made with your app’s signing secret, or it is rejected. Slack retries any event it doesn’t see acknowledged within three seconds, so Pragence queues each event, de-duplicates it, and records the reply so a retried event is never answered twice. If a reply times out after it was sent, it is not resent, because Slack has no way to post a message exactly once.
Slack answers questions only. If the assistant proposes an action there, such as sending an email, the action is refused. The setup steps are in the Slack guide.
7. Your own app
Everything above is built on the same REST API you can use directly. The Python SDK installs with pip install pragence. Your backend uses a secret key to mint short-lived tokens for your own users, then chats on their behalf, so they need no Pragence account. A publishable key is safe to put in a browser and allows anonymous read and chat only. The JavaScript client in @pragence/chat-ui follows the same token model.
There is also a read-only Knowledge query API for server-to-server use. A key used there enters one project. It does not take on the personal document permissions of the person who created it.
Choosing where to start
Start with the channel where your questions already arrive. For most teams that is the website. Phone and WhatsApp need an account with the provider first, so allow time for that setup. Test the same five questions on each channel you turn on. The answers should match, because they come from the same sources.