Meta AI Connectors: is your site ready for MCP and WebMCP?

Start with the job the assistant should finish. If someone asks Meta AI to check a site's agent readiness, a connector needs a reachable service that returns a score and evidence. If someone visits that site with a browser agent, WebMCP can expose actions inside the open page. These are two different routes to a useful result, and a site can support either or both.

MCP and WebMCP solve different parts of the request

Meta's documented connector onboarding concerns the first row. WebMCP does not automatically turn a page into a Meta AI connector. It helps a browser agent that has already opened the page. The WebMCP versus MCP comparison explains the credential and discovery differences in detail.

What the full Agent Readiness check tells you

Run the free full check on your public homepage or product page, then open the evidence under Integration and WebMCP. The scan looks for MCP discovery documents or endpoint responses, a machine-readable API description, references to document.modelContext, registered tools and tool metadata. It also checks the basics that matter to both routes: access for crawlers, useful server-delivered text, and actions an agent can understand.

Test one real task end to end

1. Choose one useful request. For example, 'Check example.com and tell me the three fixes with the most impact.' Write down the expected answer and the action that would count as success.
2. Inspect the public surface. Confirm that the API or MCP URL is reachable over HTTPS, its tools have clear names and typed arguments, and a read-only call returns current evidence rather than a generic success message.
3. Exercise awkward inputs. Try an invalid URL, a private address, a timeout and a repeated request. If a tool sends email, books or buys, require clear user intent and make the effect visible before it happens.
4. Test in the actual client. Once Meta grants preview access, onboard the endpoint there and repeat the request. Test account linking with OAuth before exposing private data. Until that test succeeds, describe the connector as a candidate, not an integration.
5. Test the browser path separately. Open the page in a WebMCP-capable browser, list its registered tools and call one with the visitor's session. A server-side MCP result says nothing about that path.

This site is an example of the two surfaces: its remote MCP endpoint offers a public score tool, while the page registers browser WebMCP tools. The scoring method explains what the scan can observe. Neither surface has been approved as a public Meta AI connector by virtue of existing.

The useful next step

Start with the full check. Fix the first access or actionability problem it finds. Then validate one MCP call and one browser tool call against a real user request. If Meta AI is your target, take that tested task into its early-access onboarding; do not build around a directory placement that Meta has not opened to the public yet.