# Providers

> The agent CLIs devpit drives, and the ones it only runs in a terminal.

devpit ships no agent CLI. It starts the ones already on your machine, and it
is honest about how far it can go with each: only Claude Code is driven end to
end. Every other CLI is one devpit can open in a terminal and recognise — and,
when it reports, show working, waiting and done like Claude Code.

| CLI | Starts with | Driven (chat, steps) | Terminal |
|---|---|---|---|
| Claude Code | `claude` | yes | yes |
| Codex | `codex` | no | yes |
| Gemini | `gemini` | no | yes |
| OpenCode | `opencode` | no | yes |
| Cursor | `cursor-agent` | no | yes |
| GitHub Copilot | `copilot` | no | yes |
| Amp | `amp` | no | yes |
| Droid | `droid` | no | yes |
| Grok | `grok` | no | yes |
| Aider | `aider` | no | yes |
| Goose | `goose` | no | yes |
| Crush | `crush` | no | yes |

"Driven" means [chat](/docs/chat), [agent steps](/docs/agent-steps) and
[session steps](/docs/session-steps): devpit holds the conversation itself and
reads the answer. Only Claude Code prints a shape devpit knows how to read, so
the rest are [terminal only](/docs/terminal-agents).

## Settings → Providers

**Installed** lists what was found on this machine, with the count detected and
a **Refresh** that looks again. **Not on this machine** lists the rest with a
link to each one's own documentation. Every row shows the line it actually
runs, and carries **Enabled** / **Disabled**, which is only about whether it is
offered in menus, and **Set default**. Opening a row edits how it starts — see
[Profiles](/docs/profiles).

**Check sign-ins and MCP servers** sets how often devpit looks again at each
account's sign-in and MCP servers while the window is on screen: every minute,
or every 5, 15 or 30 minutes, 5 by default. The chips in the footer follow a
new choice at once. **Default model** is the model a new chat starts with, for
all projects or for one; see [Settings](/docs/settings#for-one-project). A
Claude Code profile, opened, has its models, an advisor model and a subagent
model; see [Profiles](/docs/profiles#models).

<!-- Screenshot — Settings → Providers, with the installed CLIs listed. -->

## Default agent

**What a new terminal opens. A plain shell is an answer too.** One menu,
offering **No agent** — a plain shell — and every enabled CLI and profile,
each with the line it launches. If the one you chose is
no longer here, devpit says so — *"The default is not on this machine any more,
so a new terminal opens a shell."*

## Agent status hooks

A switch, on unless you turn it off: *"Shows working, waiting and done on the
island, in the sidebar and on a card, for Claude Code and the Gemini CLI run in
devpit's terminals."* The hooks travel with the launch and reach only the
agents devpit starts — nothing is written into your own configuration, so
turning this off is the whole of turning it off. See
[Claude Code](/docs/claude-code).

Gemini CLI reports from devpit's terminals with nothing to set up. devpit
points it at a defaults file of its own that adds devpit's hooks to whatever the
system already had; Gemini merges it under your settings, so your own hooks
still run and `~/.gemini/settings.json` is never written.

## Any other agent

Any other agent — Aider, OpenCode, a CI job, a script of your own — can tell
devpit what it is doing with one command, and then shows on the
[island](/docs/island) and on its card like Claude Code. `devpit-agent` is on the
`PATH` of every terminal devpit opens, and run inside one, the event belongs to
that terminal and its card:

```sh
devpit-agent hook --agent aider SessionStart
devpit-agent hook --agent aider PreToolUse --tool Edit --target src/app.rs
devpit-agent hook --agent aider Stop --said "Tests pass"
```

| Part | What it is |
|---|---|
| `--agent NAME` | up to 24 of `a-z`, `0-9` and `-`. `claude` is taken |
| the event | `SessionStart`, `UserPromptSubmit`, `PreToolUse`, `PostToolUse`, `PostToolUseFailure`, `Notification` (it waits on you), `Stop` (a turn ended), `StopFailure`, `SessionEnd` |
| `--session ID` | which session the event belongs to; without it, one per agent process |
| `--tool`, `--target` | the step it is on, as in `Edit src/app.rs` or `Run npm test` |
| `--said TEXT` | what a turn ended with, or the error |

The command always exits 0 and waits at most a second and a half: a report
that fails must not fail the agent sending it. There is no permission event.
An approval asked this way could pass for one of Claude Code's, and approving
is yours, in the agent itself.
