# WebMCP: let your browser's agent tune a Remocn component on the docs page
URL: https://remocn.dev/docs/getting-started/webmcp

Every Remocn component page exposes its Customize panel as WebMCP tools. A browser agent such as ChatGPT's built-in browser can read the props, set them in plain words, and hand you the code. No install, no key.

_Last updated: November 16, 2026_

Every component page on remocn.dev registers its **Customize** panel as
[WebMCP](https://developer.chrome.com/docs/ai/webmcp) tools. An agent built into your browser can
read the props a component takes and set them for you: "make the caption _Launch day_, half speed,
orange accent". The preview, the generated code and the page URL update as if you had moved the
controls yourself.

Nothing to install on your side, and no key. The page offers the tools; your browser's agent
decides when to call them.

| | |
|---|---|
| Where | Every docs page with a live preview and a Customize panel |
| Tools | `snapshot`, `query`, `act` |
| Sees | The Customize panel only: its controls, groups, Reset, Undo and Redo |
| Changes | Prop values, open and closed groups, Reset, Undo, Redo |
| Never does | Copy to your clipboard, share links, install anything, leave the page |

## WebMCP or the MCP server?

They do different jobs and work well together.

- **WebMCP** works on the page you are looking at. The agent tunes one component's props in the
  live preview, and you watch it happen.
- **The [MCP server](/docs/getting-started/mcp-server)** works from your editor or chat. Your agent
  searches the whole catalog, reads docs, plans a video and installs components in your project.

A common path: find a component with the MCP server, open its page, tune it with WebMCP, then take
the code into your project with **Open in** (see [below](#take-the-result-into-your-project)).

## Set it up

### ChatGPT desktop app

ChatGPT calls WebMCP tools "site tools". They work in the browser built into the ChatGPT desktop
app.

1. In ChatGPT, open **Settings → Browser → Permissions** and turn on **Enable site tools**.
2. Start a Work conversation with a model that supports site tools.
3. In the built-in browser, open a component page, for example
   [Countdown](https://remocn.dev/docs/numbers/countdown).
4. Select **Site tools** in the address bar, then **Available site tools**. You should see
   `snapshot`, `query` and `act`.

Model and workspace availability is listed in
[ChatGPT's site tools guide](https://learn.chatgpt.com/docs/webmcp).

### Chrome

Chrome has WebMCP behind a flag while the API is in an origin trial:

1. Open `chrome://flags/#enable-webmcp-testing`, set it to **Enabled** and restart Chrome.
2. Install one of the inspector extensions linked from
   [Chrome's WebMCP docs](https://developer.chrome.com/docs/ai/webmcp). They list a page's tools
   and can call them through Gemini with your own API key.
3. Open a component page. Chrome DevTools also shows the registered tools under
   **Application → WebMCP**.

### Check that the page offers the tools

Open the browser's console on a component page and run:

```js
typeof document.modelContext?.registerTool
```

`"function"` means the browser supports WebMCP and the panel has registered its tools.
`"undefined"` means this browser doesn't support WebMCP yet. The page works as usual either way.

## Try it

Start with a prompt that only reads the page:

```text
Use this page's site tools to list the Customize controls and their current values.
```

Then ask for a change in plain words:

```text
Make the final label "Launch day", the speed 0.5 and the accent orange. Switch to the light theme.
```

And step back if you don't like it:

```text
Undo the last change.
```

The agent reads the controls with `snapshot`, looks up a select's choices with `query`, and changes
values with `act`. Each change lands in the live preview at once, the Code tab shows the new props,
and the URL keeps them, so **Share link** shares exactly what you see.

## Tools

| Tool | What it does |
|---|---|
| `snapshot` | Lists the panel's controls with their labels, current values, ranges and choices. Read-only |
| `query` | Finds a control or lists a select's choices without opening it. Read-only |
| `act` | Runs steps in order: set a value, toggle a switch, open or close a group, press Reset, Undo or Redo |

A step names a control by its visible label, such as `Speed` or `Accent`. Values are checked like a
person's input: a slider stays within its range and snaps to its step, and a select takes only one
of its options.

## What the agent can't do

The agent only sees the Customize panel. The rest of the page stays yours:

- **Copy and share.** Copy code, Copy and Share link write to your clipboard, so they are hidden
  from the agent. Ask it for the values instead, or press the button yourself.
- **Playback.** The player's play, pause and seek controls aren't part of the tools.
- **Your project.** WebMCP can't install a component or touch your files. Use **Open in** or the
  [MCP server](/docs/getting-started/mcp-server) for that.

The tools run in your tab and only while it is open. The page sends nothing to Remocn; whatever
your agent reads goes to the agent you chose to use.

## Take the result into your project

When the preview looks right, use **Open in** in the Customize panel. It opens Claude, ChatGPT or
Cursor with a ready prompt, or copies it. The prompt asks your coding agent to add the component
with the props you picked, and carries three things:

- the install command, `npx shadcn@latest add @remocn/countdown`
- the code the panel generated, with every prop as you left it
- a link to the component's Markdown docs, `https://remocn.dev/docs/numbers/countdown.md`

For Countdown tuned as above, the code in the prompt is:

```tsx
import { Countdown } from "@/components/remocn/countdown";

<Countdown
  theme="light"
  accent="#ff6600"
  finalLabel="Launch day"
  caption="remocn 2.0"
  indicator={true}
  from={3}
  beat={20}
  speed={0.5}
/>
```

Going the other way, the MCP server's `preview_component` answers with a link to the docs page
with your props already set, so you can open a chat preview in the full Customize panel and keep
tuning there.

## FAQ

**Do I need an API key or an account?**
No. The tools are part of the public docs pages. Your browser's agent may need its own account, such as ChatGPT.

**Which browsers support WebMCP?**
The browser built into the ChatGPT desktop app supports it as site tools. Chrome supports it behind the `chrome://flags/#enable-webmcp-testing` flag during its origin trial. Other browsers will pick it up when they ship `document.modelContext`.

**Can the agent change anything outside the Customize panel?**
No. It sees the panel's controls, groups, Reset, Undo and Redo, and nothing else on the page.

**Does it work on every docs page?**
On every page with a live preview and a Customize panel. Guides and pages without controls register no tools.

**Can I undo what the agent did?**
Yes. Press Undo, or Cmd+Z (Ctrl+Z) while the panel has focus. An agent's change is one step, like yours.

## What's next

- [MCP server](/docs/getting-started/mcp-server): search the catalog, plan a video and install
  components from your editor
- [Agent Skill](/docs/getting-started/agent-skill): teach your agent Remocn's timing rules and
  recipes
- [Components](/docs/components): every component, each with a Customize panel to try