# Claude Code

> The CLI devpit drives end to end: chat, agent steps and session steps.

Claude Code is the one CLI devpit drives end to end. Chat, agent steps and
session steps all run through it; every other CLI it knows is
[terminal only](/docs/terminal-agents).

## What it needs

`claude` on your PATH, signed in as you would sign it in anywhere else. devpit
does not install it and does not carry a copy of it — if it is not there, the
row in **Settings → Providers** says so and links to its documentation.

## What runs through it

| Surface | What it is |
|---|---|
| [Chat](/docs/chat) | a conversation in devpit's own window, with its cost, its questions and its checklist on screen |
| [Agent steps](/docs/agent-steps) | one headless turn per card, answering in JSON, under a spending cap |
| [Session steps](/docs/session-steps) | a session started for a card, attached to the terminal when you want it |

Everything else — the [terminal](/docs/terminal), the board's command steps —
is the same whichever CLI you use, because there devpit runs a shell or a
command and reads no agent's output.

## Sessions it can go on with

devpit keeps the id Claude Code gives a conversation, so the next turn resumes
the same thread rather than starting over. That is what makes a chat a chat,
and it is also why a finished session on a card can be carried on: **Open as
chat** on the card's **Sessions** row picks the conversation up where it
stopped. Not while it is still running — two CLIs going on with one session
would each write their own continuation of it.

The chat pane also has a button that copies the session id, which is what
`claude --resume` takes if you would rather go on outside devpit.

## Remote Control

Claude Code's Remote Control makes a conversation reachable from claude.ai and
the Claude app. A chat's phone button turns it on for that chat, as
`devpit-<project>` — not in **Supervised**, which runs a process per turn. An
[orchestrator's](/docs/orchestrator-sessions) Sessions panel types
`/remote-control` into a session in one of devpit's terminals. See
[Chat](/docs/chat).

## Plan limits

[Usage](/docs/usage) shows a **Plan** block: a bar per quota window,
with how much of it is used and when it resets. It is read with the sign-in
Claude Code already saved on this machine, from the same place the CLI's own
usage command reads, at most once every five minutes — and it spends no tokens.
An installation that runs against another provider has no Claude plan limits,
and says that instead of a number.

## Hooks

devpit asks Claude Code to report what it is doing — working, waiting, done,
and the step it is on, such as *Edit invoice.ts* — so the [island](/docs/island),
a card and the sidebar can say which without guessing from the process table. The hooks are handed over on the command line, as an *additional*
settings file kept in `~/.devpit`, so what you have configured yourself is
untouched and the hooks reach only the agents devpit started. The switch is
**Agent status hooks** in [Settings → Providers](/docs/providers).

Other agents can report the same way with `devpit-agent hook`; see
[Providers](/docs/providers).

## The devpit plugin

Those hooks, and the board's tools, reach only the sessions devpit launches
itself. A session you start with your own command — a shell function, an
alias, a wrapper — gets neither. The devpit plugin closes that gap.

When a terminal is running Claude Code and the plugin is missing or out of
date, the strip under the terminal offers **Install devpit plugin** (or
**Update devpit plugin**). Nothing is installed until you press it. devpit
writes the plugin into its own folder, adds it to Claude Code as a local
marketplace, and installs it for every Claude Code installation it found. Then
run `/reload-plugins` in Claude Code.

The plugin's hooks do nothing outside a devpit terminal, and where devpit's own
launch already brought the hooks and tools, the plugin stands aside instead of
reporting twice. **×** dismisses the offer until a later build changes the
plugin.
