The Agent Filesystem Terminal lets you drive an agent filesystem by hand, using the same command set an AI agent gets through its agent_terminal tool.

Because both run through one evaluator, what you can do here and what an agent can do can't drift apart – which makes this the place to explore a volume, or to exercise the command layer before any agent is wired to it.

Access
Mounting
Commands panel
Running commands
Output
The session is not saved
Compared to an agent

Access

Click Setup » Developers » Agent Filesystem Terminal. This tool is limited to administrators.

Mounting

Nothing is mounted when you arrive. Add agent filesystem records from the chooser at the top right, and each appears as a chip in the Mounts panel. You can mount several at once; mounting the same volume twice does nothing.

Each volume mounts at its own name. Unlike the mounts: block an automation composes for an agent, you can't remap a volume to a different path here.

Every mount arrives read-only. Switch a chip to rw to allow write, append, edit, rm, and writing with copy.

The ro/rw switch is a workflow safeguard, not a security boundary -- it protects you from a mistyped rm, not from a determined operator. Access control for this screen is the administrator requirement itself.

/tmp is always writable, since it's where oversized command output is spilled.

Commands panel

Beside the mounts, a Commands panel turns each cerb command namespace on or off for this session – the same command line an automation gives an agent through its terminal: block.

With records enabled you can run cerb records types, cerb records filters <type>, and cerb records fields <type> here exactly as an agent would, which is the quickest way to see what an agent will be told about your installation.

With no namespace enabled the cerb command isn't merely hidden – it doesn't exist, and typing it says so.

Running commands

Type help to list the commands, or help <command> for one of them. Tab completes both command names and paths, and the up and down arrows recall previous commands along with their payload.

The prompt shows the working directory, which persists between commands the same way it does for an agent.

Payload and Find

Two boxes below the command line carry the things that don't belong on it.

Payload does three jobs depending on the command:

With	Payload holds
`write`, `append`	The file's content
`edit`	The replacement text
Anything else	A multi-line Twig template to evaluate

Find holds the search text for edit – the snippet to be replaced, which must match exactly once.

Keeping these out of band means a script or a block of file content never has to be escaped onto a command line.

Output

The output pane is a fixed height and scrolls with new output pinned to the bottom, so a long result scrolls the command that produced it out of view.

There's no paging. For output larger than the pane, the intended tools are the same ones an agent uses: pipe through a filter chain to narrow it, or read from the /tmp file that oversized output spills into.

search widgets | results|filter(r => r.hits > 3)|column("path")

Clear wipes the scrollback only. Your mounts, working directory, and /tmp contents survive it, the way clear behaves in a shell.

If the pane flashes, that's the terminal bell: Tab completion had either nothing to offer or too many choices to pick from.

The session is not saved

Everything here lives in your browser. The mount set, the working directory, and the /tmp scratch area are held in the page, and the server keeps no state between commands. Reloading -- or navigating away and back -- returns you to an empty terminal with nothing mounted, without warning. Nothing you do here is lost from the volumes themselves; it's the session setup that goes.

Compared to an agent

An agent gets the same commands, the same evaluator, and the same persistent working directory. Two things differ:

An agent's mounts come from the automation's mounts: block, resolved on the server. Here you compose the mount set by hand.
The Payload and Find boxes are specific to this screen. An agent passes the same values as part of its tool call.
The cerb namespaces an agent may run come from its automation's terminal: block. Here you enable them yourself, from the Commands panel.