Documentation Last reviewed 2026-08-08

Integrations and models

What actually connects, in which direction, and which model providers are real.

An icon is not an integration. This page lists what is connected today and what is not, in the same table, because the second half is the part that is usually missing.

Channels

A channel is how a message reaches an agent and how the reply goes back.

ChannelStateDirectionNotes
EmailConnectedReads and sendsStandard IMAP and SMTP. Works with any provider that speaks them, including Google and Microsoft
TelegramConnectedReads and sends
SlackConnectedReads and sendsThe installer does not wire it. We do, after install
Web chatConnectedReads and sendsThe panel, and a widget on your site if you publish one
VoiceConnectedReads and sendsAnswers a call, delegates, and replies on the same call
WhatsAppOn requestReads and sendsThe channel layer supports it; the adapter is built when a customer needs it
SMSOn requestReads and sendsAs above

All the working channels open the connection outward. None of them requires a port to be opened toward your machine.

Everything else

ConnectedNothing beyond the channels above
Partial, not usable yetGoogle Calendar, Google Drive
Not built, available on requestSheets, Meet, Notion, HubSpot, QuickBooks, Stripe, Zapier

Google Calendar and Drive have connectors written and neither is reachable by an agent yet. Drive access, when it is finished, is scoped to files the node itself created; the other Google connectors ask for their own scope, which you approve in your own Google account.

Two things follow from this table and are worth saying plainly. First, the list is short. Second, a new connector is a small piece of work. A tool is a row in the registry with a body behind it. If you need one, the question is when, not whether.

Models

One gateway carries every model call, so changing provider is a configuration change. Wired today:

ProviderReaches
AnthropicClaude
GoogleGemini
OpenRouterMost of the rest, including OpenAI

There is no local model. The only model that runs on your machine is the small one used to index your own text for search. If running inference locally matters to you, it matters now, before a pilot and not during one.

What self-hosted means here

The word gets used loosely, so it is defined once and this is the definition the rest of the site means:

Your data, your configuration, the decision logic and the audit log are on your machine, and
that machine needs no inbound connection from the internet. The model call goes out. Nothing
else does.

Paying for models

Two ways, decided during setup.

Managed. We hold the provider accounts and you get one bill.

Bring your own. You plug in your own keys and pay the providers directly. Independent of us from that point on.

There is no subscription for the software either way.

Reviewed 2026-08-08. Older than 90 days counts as broken. Something here wrong? Tell me