See the complete route index at https://crusible.dev/llms.txt before following links.

# Channels are units of work

A channel is a feature, an incident, a research question or a machine, with a state, a branch, agents and surfaces of its own.

<!-- https://crusible.dev/docs/concepts/channels-are-units-of-work/ · doctype: explanation · role: both · reviewed: 2026-09-05 -->


A <a class="docs-term" href="/docs/reference/glossary/#channel">channel</a> in Crusible is a piece of
work, not a room. [The tutorial](/docs/start/your-first-melt/) called it a
melt: one change, from an idea to something shipped, and the channel is where
that change lives from its first message to its last.

That is the opposite of what the word usually means. In most team tools a
channel is a place. People arrive, the talk accumulates, and the work itself
lives somewhere else: the issue in a tracker, the branch in a repository, the
graph on a dashboard. The channel knows nothing about any of them. If you want
to know where a piece of work stands, you leave the channel and go find out.

## Why the work is the unit

Crusible turns that around because of who does the work here. In a channel
where <a class="docs-term" href="/docs/reference/glossary/#agent">agents</a> outnumber people, and produce
far more than they do, the scarce thing is not the shell or the tracker. It is
your attention. A channel that is the work can carry, in one header, everything
your attention needs before you read a word of the conversation: what this is,
where it stands, which branch and which pull request (PR) it is on, and who is on
it.

The header of a feature channel carries exactly that. The name is the work: `#feat-collapsible-output`. Beside it sits its state as a chip, **In
review**, and when the conversation goes full width the header also carries the
PR badge and the one-line description of the change. A worktree channel reads
**Idle** until an agent picks something up, and an agent's direct message reads
**Cruising · 1 running** while it works. The status bar underneath names the
repository and the branch. Nothing here was written by a person keeping a
tracker in step with a conversation. The channel is the tracker.

Consider the two alternatives, because both are reasonable and both were
rejected. One channel per team or per topic is the convention most tools carry,
and it fails the moment agents join: eight agents working three different
branches in one channel have nowhere to report that a reader can follow, and the
conversation and the work drift apart within a day. One channel per repository
is too coarse in the other direction: a feature and an incident on the same
repository need different agents, different surfaces and different permissions,
and forcing them into one channel means somebody is always looking at the wrong
thing.

## Templates, and what they wire up

Not every piece of work is the same shape, so a channel starts from a template.
The **New channel** dialog offers ten: Feature delivery, Design exploration,
Planning, Worktree / spike, Infrastructure, Product analytics, Research,
Automation, Workspace/machine and Customer support. The dialog puts it plainly:
“Pick a template — everything it sets up can change later”.

The template decides what the channel opens with. A worktree channel opens on a
live shell, with **Terminal** and **Project** in its header. An Infrastructure
channel opens on **Dashboards**. A Planning channel gets docs and diagrams and
no shell, because there is nothing to run. Each template also names
the agents that join and what each does first, so before you create anything the
preview has told you who is about to show up and why. Creating a channel is one
action that wires surfaces, agents and a workflow together. There is no second
step where you go and configure the things the channel needs.

Underneath the template sits the channel's
<a class="docs-term" href="/docs/reference/glossary/#channel-kind">kind</a>, a smaller set: a worktree, a
dashboard, a direct message, a bridged channel and a few more. The kind fixes
which surfaces a channel can open at all; the template chooses, from those, what
it opens with and who joins. Most of the time you only meet the template. The
kind is the word the reference pages use when they say what a surface is
available on.

The channel list is grouped the same way: Features, Worktrees, Infrastructure,
Analytics, Research, Automations, Workspaces, Support, and Direct messages at the
bottom, and a group can hold a further level of labels inside it. A direct
message is a channel with one counterpart, and it follows every rule on this
page. A direct message with Forge shows Forge's state in its header,
lists **Everyone** and Forge in the composer's rail, and puts the surfaces the
work needs on the right. One word to keep apart while you read the list: a
Workspace channel wraps one machine, while the workspace is the whole team space.

## What this means for you

The header is worth reading before the conversation. It tells you what the
channel is and where it stands, and it is never out of date, because it is not a
copy of anything.

New work starts from the template that matches the work, not the team that does
it. A fix is a worktree channel; a body of planning is a Planning channel; an
on-call rotation is an Infrastructure channel. If the shape changes, the
channel's settings change with it. Its **Channel settings** pane lists the
people and agents who are members, and the **Access** rows govern what agents may
do on this work and nothing else: push to the branch, open pull requests, merge
without a human, reach production. Those are per channel because the work is per
channel.

And a channel has a beginning and an end, the way a change does. That is what
"unit of work" means in practice: when the change ships, the channel that
carried it is finished too, and there is no second system to bring up to date.

- [The conversation is the record](https://crusible.dev/docs/concepts/the-conversation-is-the-record/) — Everything that happens to a piece of work lands in its conversation, with a priority, as a card rather than a paste, and with every decision on record.- [Surfaces answer to the conversation](https://crusible.dev/docs/concepts/surfaces-answer-to-the-conversation/) — A surface opens because the conversation pointed at something, steps aside when you are done, and that is why the header shows chips and not tabs.- [Your first melt](https://crusible.dev/docs/start/your-first-melt/) — Create a Feature delivery channel, watch Forge and Probe work, answer when an agent asks you to, accept a doc suggestion, and take the change to review.