> ## Documentation Index
> Fetch the complete documentation index at: https://docs.canvasview.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# The canvas's agent

> Every canvas has its own agent. Editors ask it to change or explain the canvas from the canvas page.

Every canvas has its own agent. Editors and admins of the canvas can ask it to change the canvas or explain it, right on the canvas page. It works beside the agent that built the canvas, not instead of it.

## Ask the agent

1. Open the canvas and select **Agent** in the bar.
2. Write what you want, such as "Add a column for the due date" or "What does this chart show?". **Enter** sends, **Shift+Enter** starts a new line.
3. Select **Send**.

The reply streams in as it is written, with a line for each step the agent takes, such as **Reading the canvas…** or **Changing the canvas**. Select **Stop** to end a run. What it changed up to then is saved.

In the **Components** view, the **+** tab also offers **Ask the agent for a new component**.

Everyone on the canvas sees who is asking the agent. A prompt can be up to 4,000 characters. Up to five prompts can wait their turn, marked **Waiting**.

## Who can use it

Only people with the editor or admin role on the canvas see the **Agent** button. See [Canvas roles](https://docs.canvasview.dev/account/sharing-and-access.md#canvas-roles). The agent works on active canvases only.

## What it can do

The agent can read the canvas, its answers, shared state and change log, change and arrange components, write shared state, comment, and run the canvas's tools. It cannot publish new tools.

One editor changes a canvas at a time. While the agent works, editors see **The canvas's agent is editing this canvas.** in the bar. While someone else is editing, the panel names them and takes no prompt until they finish. Each change the agent makes is recorded in the canvas's change log as **Canvas agent**, with who asked. Agents read the log with `list_canvas_changes`.

A comment with **Ask the agent to change the canvas** goes to the agent that built the canvas, not to this one. See [Comments and change requests](https://docs.canvasview.dev/people/comments-and-change-requests.md).

## Limits for one request

A run stops after 12 steps, 2 minutes, or \$0.25 of model use. The agent says which limit it reached. What it changed is saved, and you can ask again.

## Cost

Runs are paid from your organization's credits:

* the model use of each run, at most \$0.25;
* each change it makes to the canvas, priced like any other update. A change that only refreshes data is free. See [Credits and billing](https://docs.canvasview.dev/account/credits-and-billing.md).

Runs count toward the organization's daily spending cap, but not toward any key's cap. They show in **Daily usage** on the Billing page. When credits run out or the cap is reached, the run is refused with the same errors as other paid requests.

## What is logged

CanvasView keeps a record of each run with the canvas: who asked, the model, how much it used and cost, and what it did. A full trace of each run, including what was asked and the canvas content the agent read, is also logged with Cloudflare, which runs the model, and with PostHog. Cookie settings does not change this. See [Security and data handling](https://docs.canvasview.dev/trust/security.md#analytics).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.