Channels

A channel is how you talk to your assistant. The assistant is the same everywhere: same name, same memory, same skills. What changes is where you are talking and what that surface can do.

Linux desktop is first. 0.1 is a packed AppImage or deb on your machine. There is no Cadence Cloud sign-in, and no WorkOS product page. You hatch the assistant locally and bring your own model keys.

Linux desktop

The flagship 0.1 experience. Install the Linux AppImage or deb, hatch an assistant, name it yourself. Chat, rich surfaces, and native permission prompts live here.

Chat: messages, cards, tables, and other UI surfaces
Approvals: native in-app prompts when a tool needs your go-ahead
Host file and shell: with your confirmation, the assistant can read files and run host_bash on this machine (see The permissions model)
Voice input: where the desktop client supports it

Host tools are confirm-gated. They are not unrestricted hands on the machine, and they are not a remote-admin path onto a house server.

CLI

Packaged Linux installs place cadence on your PATH (often ~/.local/bin/cadence). Open a new terminal after first launch. The CLI talks to the same local assistant as the desktop app: streaming replies, permission prompts, and sandbox skills.

Useful for scripting, or if you prefer a terminal. Live state (model, skills, channels, trust) is what the CLI reports. See Glossary.

Local web

The desktop shell loads the web app against the assistant running on this computer. That is not a hosted Cadence Cloud URL. The browser does not become a cloud home: workspace and sandbox stay on disk.

Connect a Telegram bot and message the assistant from anywhere. Typical setup: create a bot with BotFather, store the token through the credential prompt (never paste secrets in chat), and let the assistant register the webhook.

Text, images, documents, and inline buttons for approvals. The assistant can also reach you here with notifications. First contact on a new channel goes through guardian verification so the assistant knows it is you.

Slack

Connect a Slack workspace. The assistant can reply to mentions, hold threads, send Block Kit messages, and handle approvals with buttons. Slack can carry per-channel permission cells (for example, coding tools in an engineering channel, tighter limits in a general one).

Email

Your inbox

Connect Gmail (or another supported mail provider) so the assistant can read, search, draft, and organize mail. Sending stays draft-first: nothing goes out without your approval.

The assistant's own address

An assistant-owned inbox is optional and feature-flagged. It is not a Cadence Cloud mailbox. If you enable it, inbound mail to that address becomes a conversation thread.

Phone and SMS

Phone and SMS use Twilio. You bring your own Twilio account and number. Inbound calls can reach the assistant; outbound calls are for cases like guardian verification. SMS is a delivery channel, not a marketing product. Phone is the most constrained surface: the assistant cannot start a random call the way it can send a Telegram message.

Windows, macOS, iOS, Android

Those clients exist in the open-source tree. They are not the 0.1 ship target. Linux is first. Do not expect an App Store listing or a Cadence Cloud iPhone login in this release.

Channels vs clients

Channels are transports: the Cadence client channel, Telegram, Slack, email, phone. Every message has a channel.

Clients (also called interfaces) are the specific apps on that Cadence channel: Linux desktop, CLI, and later web, Windows, macOS, iOS. Interactive clients can show native permission prompts. Telegram and Slack route approvals through buttons or a text reply.

What works where

Capability	Linux desktop	CLI	Telegram / Slack	Email	Phone
Chat	Yes	Yes	Yes	Yes	Voice
Host file / `host_bash`	Confirm-gated	No	No	No	No
Approvals	Native prompt	Native prompt	Buttons or text	Plain text	Limited
Notifications	Yes	Yes	Yes	Optional	No

The assistant adapts output to the channel. An interactive card on desktop becomes markdown on Telegram or spoken words on a call.

The guardian

You are the guardian: the person who hatched and owns the assistant. Each assistant has exactly one guardian. The guardian has full access to memory, workspace, tools, and credentials, and is the only one who can grant sensitive actions.

Verify who is talking. The first time you message on a new channel, a short challenge links that identity (Telegram user, Slack handle, phone, email) to you.
Route approvals. Sensitive tool calls ask the guardian. On desktop and CLI the prompt is inline. On Telegram or Slack you reply there. Channel approvals are one-time grants, not standing host access.
Gate memory. Long-term memory is extracted from guardian conversations. Messages from contacts or unverified parties stay in that thread; they do not rewrite what the assistant knows about you.

Details: The permissions model.

Setting up channels

Desktop and CLI need no cloud account: install, hatch, talk. Other channels are configured through the assistant. You can say “set up Telegram” or “connect Slack.” The assistant walks the flow. API keys and tokens go through the credential prompt, never into the chat.

Some channels need an external account (Twilio, BotFather). The assistant will say so. Check live delivery with cadence channels list (or assistant channels list in the assistant shell).

See also Scheduling, Local hosting, and The workspace.