The [WebMCP draft dated September 4, 2026](https://webmachinelearning.github.io/webmcp/#sotd) is a Community Group report. Its publisher is the Web Machine Learning Community Group at W3C, the web standards organization. That draft is neither a W3C Standard nor a Standards Track specification.

The proposal lets a website describe actions directly to an agent. That gives the agent another way to interact alongside the page's visible controls.

## What does a page expose?

A page exposes named tools with descriptions and structured inputs. Its application code implements the requested behavior.

The [WebMCP draft](https://webmachinelearning.github.io/webmcp/#intro) describes JavaScript tools and a declarative form interface. A tool can describe an operation that otherwise requires the agent to infer how to use the page.

For example, an appointment page could expose a search operation with location and date inputs. The agent supplies those values instead of reconstructing the search from labels and controls.

## How is this different from a remote MCP server?

WebMCP exposes functionality from a running web page. A remote MCP integration connects a client to a server outside that page.

The [WebMCP explainer](https://github.com/webmachinelearning/webmcp#backend-integrations-vs-in-browser-webmcp-tools) describes the distinction in terms of shared browser context. Page tools can reuse the web application's code. They can participate in a workflow the user can see.

A backend integration instead connects the agent's application to a service endpoint. The service can expose operations independently of a particular open page.

| Integration | Where the capability is exposed | Example                                                 |
| ----------- | ------------------------------- | ------------------------------------------------------- |
| WebMCP      | A running web page              | Search appointments using the page's application code   |
| Remote MCP  | A server endpoint               | Invoke a messaging operation through a connected client |

A site can use both approaches. A page tool may call a backend, so browser exposure does not mean every action happens locally.

## Does a tool description make an action safe?

A description tells the agent what a tool claims to do. The application still needs to enforce permissions and validate the request.

The draft's [security and privacy considerations](https://webmachinelearning.github.io/webmcp/#security-privacy) address misleading tool metadata, injected instructions and unintended actions. Structured inputs do not establish that a requested action matches the user's intent.

For a booking workflow, finding available appointments and confirming a booking have different consequences. The application should preserve that distinction when it exposes tools.

A publisher's tool registration also does not establish that every browser can invoke it. [Chrome's WebMCP documentation](https://developer.chrome.com/docs/ai/webmcp) describes its own implementation and limitations.

## How do I connect an agent to Bird for messaging?

You can connect a compatible client to [Bird's hosted MCP server](/docs/ai/mcp-server) and complete the browser sign-in.

That is a remote MCP connection. It gives the client access to Bird's documented tool interface rather than declaring tools in an open Bird web page.

After sending, retain the message identifier and check its later status. The [existing MCP answer](/explained/platform/what-is-an-mcp-server-and-how-does-an-agent-send-messages) covers the connection and messaging workflow.

## Related resources

- [Send your first email](/docs/get-started/send-your-first-email) (docs)
- [Should I use a Bird SDK or call the API directly?](/explained/platform/should-i-use-an-sdk-or-call-the-api-directly) (answer)
- [Build your first integration](/learn/paths/integration) (course)

[Get an implementation brief](/learn/workspace?topic=api-basics)