Working with an AI agent

Plicine can host an AI agent of your own, such as Claude Code, so you can ask for a change while looking at the document it's about: "tighten this introduction", "add a row for travel to the quote", "make me a block that shows our opening hours".

The agent isn't part of Plicine. It's a program on your computer, under your own account with whoever makes it, and your conversation goes to them, not to us.

What you need first

One of these agents, installed and signed in on your computer:

• Claude Code
• Kiro
• Gemini CLI
• Codex

Installing one is done in a terminal, following its maker's instructions. If that isn't your kind of thing, ask a technical colleague; it's a one-off. Plicine finds whichever ones you have.

Opening the palette

The palette is the bar at the bottom of the document. Click it, or press Cmd+J (Ctrl+J on Windows) from anywhere. It's also View > Agent Palette, and the row at the foot of the sidebar shows which agent is picked and what it's doing. Esc folds the palette away.

The first time, pick your agent from the list. If it isn't signed in, the palette says so and offers a button that puts it right.

Asking for something

Type what you want. Enter sends and Shift+Enter starts a new line.

The document you're reading goes along with your message, so "this paragraph" and "the table above" mean what you'd expect. Remove its label under the box if you'd rather it didn't.

Select some text first and the agent sees that too. A second label appears under the box, such as Selected text: 3 paragraphs, and the agent is told which lines you selected and what they say. "Shorten this" then means the selection, not the whole document. The text stays selected on the page while you type in the palette. Click the cross on the label to send a message without it. It works in both layouts, and for text selected inside a paragraph you're editing.

Plicine also gives the agent a briefing about your vault at the start of a conversation: how documents and blocks work, which blocks you have, and which files to leave alone.

While the agent works, the palette shows its plan, each file it reads or changes, and its reply as it arrives. When it's thinking, a Thinking… line shimmers at the bottom with the latest thing it's thinking about beside it; click it to read the whole thought. Claude sends a summary of its thinking rather than every word. The document updates as the agent saves. Anything you send in the meantime waits its turn. Fold the palette and a strip keeps the current step in view, with a Stop button.

Type / in the box for commands: Plicine's own (answer the document's comments, export a PDF, switch the layout, start a new conversation, stop the agent) and any the agent has. Answering the comments is also a pill in the bar itself whenever the document has some waiting on an answer from you; Comments says what the agent does with them.

When the agent asks permission

Agents ask before doing some things, depending on the mode they're in. When yours asks, the palette shows what it wants to do and where, with the before and after for a change to a file, and waits. Pick one of the answers it offers, such as allowing it once, allowing it for the rest of the session, or saying no. The menus at the top of the palette change the agent's mode and, for agents that offer a choice, which model it uses.

Turning agents off

If you don't want an AI agent in a vault, open the vault's settings and turn AI agents off. The bar under your documents goes, along with its row at the foot of the sidebar and the note that points it out the first time. The shortcut does nothing. An agent that was running is stopped, and Plicine refuses to start another one for that vault, so nothing carries on out of sight.

The setting belongs to the vault, so you can have agents in one folder and not in another. Turn it on again and everything comes back where it was, with no need to reopen the app.

It has no say over an agent you run yourself from a terminal, and it leaves Write AGENTS.md alone.

Worth knowing

• An agent acts as you. It can do to your files what you can. In a folder other people write to, keep it in a mode that asks before changing anything, because a document could contain text written to steer it.
• A block the agent makes still waits for your yes before it runs, like any other new block.
• The conversation lives in the open app. It's gone when you close Plicine.
• There's one agent and one conversation for each open vault.

For technical readers: Building with an AI agent covers running an agent from a terminal, the tools Plicine gives it and the AGENTS.md briefing file.