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

# WebMCP

> The assistant in your browser, using the app as you.

<Note>
  **Platform features are currently in invite-only beta.** You can
  [sign up here](https://www.cordango.com/beta/) if you want to be considered for one of our next
  phases. The open foundation, meaning the CLI, the compiler and the standalone
  generator, is not part of the beta and is available now.
</Note>

MCP is for a client somewhere else, holding a key. **WebMCP** is for the assistant built into the
browser you are already signed in to. Cordango registers tools on each app's pages, and an agent in
that browser can call them.

<Note>
  WebMCP is a proposal at the W3C Web Machine Learning Community Group, in an origin trial in Chrome
  and Edge. Whether it is available is your browser's decision, not ours. Each app's
  [Connect page](/mcp/connect) says plainly whether this browser can use it.
</Note>

## What makes it different

There is **no endpoint and no credential**. The tools run inside the tab, through the same session
you already have, so an assistant reaches exactly what you could reach by clicking and nothing at
all when you sign out. Nothing you do this way leaves the browser except the ordinary API calls the
app was going to make anyway.

That is why it is on by default and why the tool set is not the same as the MCP one.

## The tools

| Tool                | What it does                                                            |
| ------------------- | ----------------------------------------------------------------------- |
| `describe_app`      | The open app: its entities, their fields, and what you may do with each |
| `list_records`      | A page of records, with filters and sorting                             |
| `get_record`        | One record by id                                                        |
| `aggregate_records` | Count or total, computed by the server                                  |
| `create_record`     | Add one                                                                 |
| `update_record`     | Change the fields it names                                              |
| `run_command`       | Run one of the app's own actions                                        |
| `open_page`         | Show one of the app's pages                                             |
| `open_record`       | Show one record                                                         |

The last two are the ones a browser can do and a server cannot. An assistant that can answer "where
do I see this" by putting it on screen is worth more here than another way to read rows.

**There is no `delete_record`.** Deleting is irreversible and a misread sentence is a poor reason to
lose a record, so the browser tools navigate you to the record instead and let you decide.

## What asks before it acts

Every write is marked *consequential*, which is the browser's cue to confirm with you before running
it. Every tool that returns record data is marked as carrying content the app's own users wrote, so
an agent treats it as data rather than as instructions.

Results are kept small on purpose — ten records by default, twenty-five at most. An answer much over
a page tends to be dropped by the agent rather than truncated, and a dropped answer looks like a tool
that does not work.

## Turning it off

It is a workspace setting, like MCP: an administrator can switch `webmcp` off and the app stops
registering anything. A browser that does not support WebMCP registers nothing either way, and the
application behaves exactly as it always has.