Documentation
MCP, WebMCP, or both?
MCP connects an AI app to a server. WebMCP attaches tools to the page a person has open. Use both when the same reviewed capability should work in both places.
Choose how people should connect
| Choose | When it fits | What you provide | Keep in mind |
|---|---|---|---|
| MCP | People should use your tools from an AI app, coding client, or other connected service without keeping your website open. | An MCP server that publishes tools and accepts calls over a supported transport. | The AI app controls its connection flow and interface. Authentication, authorization, and approval still need to match each tool. |
| WebMCP | An agent should use tools attached to the current page and work with that page's live context and interface. | Page code that registers tools through the browser's WebMCP interface. | WebMCP is a developing browser specification. Compatible browser and client support is not universal. |
| Both | The same capability is useful from a connected AI app and from the page where a person is working. | One reviewed tool definition published through MCP and WebMCP with the same inputs, limits, and permissions. | Test each connection separately. A working MCP endpoint does not show that WebMCP is present on the page, or the reverse. |
The MCP specification defines the client-and-server protocol. The WebMCP draft defines the page interface. Discovery files can help an agent find either connection, but retrieving a discovery file does not show that a tool was available or called.
If you offer the same tool through MCP and WebMCP, keep its name, inputs, limits, read-only or write status, and approval rules consistent. People should not get a different level of access simply because they connected another way.
How to read test results
Each label says what happened in the test. One result should not be described as another, and a tool's own success response does not confirm by itself that the requested change happened.
| Label | What it means | What it does not show |
|---|---|---|
| Native WebMCP | The browser exposed the draft document.modelContext interface and the page registered tools on it. | Universal browser support, successful tool execution, or the resulting state. |
| Extension or client compatibility | A browser extension or AI app provided its own compatible way to reach the page's tools. | That the browser provided the WebMCP interface natively. |
| Test compatibility layer | Test code provided a missing interface or stood in for a real tool call. | Native browser support or execution by an outside agent. |
| Repeatable code test | The same fixed inputs were sent directly to the code, without a model choosing the tool. | That an agent chose the right tool or completed a user's task. |
| Outside agent used the tool | An agent or AI app running outside the website found and called the published tool during a dated test. | That every agent is compatible or that the reported result was checked another way. |
| Result checked separately | The website or system was checked another way instead of relying only on the tool's response. | That every surrounding workflow will behave the same way in the future. |
A dated Open for Agents comparison found tools through Chromium's experimental Native WebMCP interface, while test-only compatibility code supplied the Kitesurf path. That second result used a Test compatibility layer. The comparison explicitly did not execute a tool, so an outside agent did not use the tool and no resulting state was checked separately. Read the comparison notes.
The Microsoft Agent Framework examples are a Repeatable code test: Python and .NET called representative MCP tools directly without a model. They show that those tested clients could find and call the server tools, not how a model would behave or whether every client is compatible. Read the interoperability guide.
Connect the hosted Website Visibility Check with MCP
Open for Agents also runs a hosted MCP server for its read-only Website Visibility Check. A compatible AI app can use it without leaving the conversation. This is a remote MCP service; it is not WebMCP attached to the website being checked.
Server address
https://gateway.openforagents.com/mcpAdd this address where your AI app accepts a custom MCP server. The app should discover the Open for Agents sign-in and approval flow automatically. Support and menu names vary between AI apps.
Opening it directly in a browser returns a JSON message about a missing Authorization header. That is expected. A compatible AI app uses the address to start secure sign-in and ask for your approval.
The server name, version and endpoint are also published in the official MCP Registry. You can connect through the Smithery listing, and Glama has imported the same record into its MCP connector directory.
What you approve
After signing in with a secure email link, Open for Agents shows the requesting app and the access it wants. Approving the connection allows that app to call one read-only tool: check_website_visibility.
The tool can:
- Request selected public HTTPS pages and files from a website you name.
- Return a visibility report to the connected AI app.
- Save up to 100 recent reports to your account for no longer than 90 days.
- Include the previous saved result for the same website when one exists.
What it cannot do
- Log in to or change the website being checked.
- Bypass access controls or retrieve private pages.
- Run WordPress tools, shopping-cart changes, Trusted Actions, or the visitor Assistant.
- Guarantee search inclusion, AI citations, recommendations, or compatibility with every agent.
How to ask
Try a direct request such as:
Check example.com with Open for Agents and tell me which public visibility issues I should address first.
Privacy and removal
Open for Agents uses Supabase to handle sign-in and connection approval. It stores which AI app you connected, the website address you checked, the resulting report, and a fingerprint used to compare results. It does not keep copies of page content, access tokens, client secrets, or website login details in report history. See the Privacy Policy for retention and request rights.
You can deny a new connection during approval. To request account-history deletion or help revoking a connection, email hello@openforagents.com.
Use the browser version instead
The public Website Visibility Check remains available without an account and stores comparison history only in that browser.