In-app agent

The in-app agent lives in the workbench and helps you edit product context under your control. It can select elements on the canvas, take comments on specific parts of a screen, run across multiple sessions, and operate under different permission modes. To use it, open a project and make sure you have AI turns remaining on your plan.

How it differs from a coding agent

The in-app agent and the coding agents you connect over the CLI do different jobs. It's worth keeping them straight:

	In-app agent	Coding agent (CLI)
Where it runs	The OriginAI workbench	Your IDE, Claude Code, Cursor, and so on
What it edits	Your specs, in the workspace	Your code, from a release
How usage is metered	Daily AI turns	An access token; no turn meter
After the first release	Keeps editing the workspace	Stages a change request

A basic turn

Open the agent panel (the right column on desktop).
Describe the change you want — a screen, a rule, a flow, or some copy.
Review the tool results and proposed file edits in the transcript.
Iterate in the same session, or open History to revisit earlier ones.
When the specs look right, publish a release.

Selecting an element on the canvas

By default the canvas is in pick mode, so clicking a rendered element attaches it as context for your next message.

Open a screen file so the canvas shows it.
Click the element you care about — a layout region, a control, or a block of text.
Confirm the composer shows a Selected chip.
Type what you want changed. The placeholder becomes "describe the change for the selected element…".
Send. The agent receives both your message and the selected element's underlying spec and path.

A few tips:

If your next request is global, clear the selection first (click empty canvas or dismiss the chip).
Prefer a selection plus a short instruction over pasting a large block of RPML when you mean "this specific control".
Cmd/Ctrl-clicking on the canvas follows navigation links (to), rather than selecting an element — use a normal click to attach a selection.

Commenting on elements

You can attach short comments to a selected element before sending your turn.

With an element selected, use the comment field ("Comment on this element…").
Commit the comment to add it to the stack — the chips appear above the composer.
Optionally add a free-text message. You can also send comments on their own.
Edit or remove chips before sending if the stack isn't right.
When you retry a turn, comments can be reattached with the same text.

Comments travel with your message, so the model sees labeled notes pinned to specific parts of the UI rather than only your free-text prompt.

Sessions and history

Control	What it does
New session	Starts a clean transcript without deleting your old sessions.
History	Lists your sessions so you can switch back to an earlier conversation.
Empty state	Surfaces a couple of recent sessions for quick resume.

Start a new session when the topic changes — say, moving from the onboarding flow to billing copy — so the agent's context stays focused. History is for resuming unfinished work, not for versioning your product; releases handle versioning.

Permission modes

The agent can pause for confirmation before high-impact steps. You choose how cautious it is:

Mode	Behavior
Auto	Confirms deletes and questions before running them.
Edit	Auto-approves deletes, but still asks questions.
Bypass	Auto-approves everything and runs fully autonomously.

Stay on Auto until you trust how the agent behaves in your project. Switch to Bypass only for long, well-scoped edits you're actively watching. Note that permission mode does not bypass your plan's turn limits, nor the post-release change-request path for coding agents.

You're doing it right when…

The specs on disk match what you asked for.
You know whether you have turns left for the day.
You can attach a canvas selection (and comments) without pasting raw RPML.
You know which permission mode is active before starting a multi-step edit.

Troubleshooting

Issue	What's happening
The composer is disabled, or you see a limit toast	Your daily turns are used up — wait for the UTC reset or upgrade.
The agent contradicts a release	Your workspace may be ahead of your last publish — publish again so agents see the change.
The selection chip is missing	Make sure a screen is open, the canvas is in pick mode, and you clicked a concrete element.
Comments send on their own	This is supported when the stack isn't empty, even if the main input is blank.
Unexpected deletes	Check your permission mode — Bypass won't ask.
A coding agent still can't write the workspace	Expected after the first release — the write stages a change request.

In-app agent

On this page