WebMCP lets a web app describe actions as structured browser tools. In Toolkit, a compatible browser agent can work with the signed-in dashboard rather than relying only on button positions. Toolkit checks inputs, uses normal authenticated APIs and asks the person using the page to approve consequential actions.
WebMCP remains an evolving browser API, not universal functionality in every browser or AI assistant. The current draft and Chrome's preview announcement describe the underlying initiative. The rest of this guide describes Toolkit's implementation and its limits.
What tools are available?
The app declares ten tools. Search the library, list workspaces and inspect a source are read-oriented operations. Creating a workspace, capturing a URL, importing URL batches, tagging sources, preparing a briefing, exporting a workspace and sending a source to a notebook use the same application workflows as the interface. Individual operations can still be unavailable because of account limits or feature settings.
See the live tool catalog and playground for exact inputs and examples. A tool being listed is not a promise that every plan can execute every action. It is also not permission for an agent to read unrelated accounts or bypass the server's ownership checks.
A sensible first task
Start with a read-only request: find sources about a topic in your current library and list the relevant workspace. Inspect the results before requesting a change. If you then ask the agent to capture a public article, specify the destination and review the confirmation prompt before approving it.
For a notebook handoff, keep the extension installed and the intended notebook open. A tool result may represent queued work rather than completed delivery. Verify the source in the notebook before using it as evidence. The import recovery checklist explains how to distinguish the stages.
What approval does and does not mean
Approval covers the action described in the prompt. It is not blanket consent for future changes. Declining an action returns a refusal instead of running the write. Source text is untrusted material: instructions found inside an article should not be treated as instructions from the account owner.
Read-only access still matters for privacy. A tool that returns source text makes that text available to the agent handling your request. Only use an agent you trust with the selected content. Review the privacy policy and avoid sending confidential research to a service that is not approved for it.
Is this a remote MCP server?
No. This integration runs in the web page and depends on a compatible browser environment and an active signed-in session. It does not give you a remote server URL to paste into a desktop assistant, run an unattended background worker or provide a human support chat. The ordinary dashboard remains available when the browser does not expose WebMCP.
Does WebMCP improve search rankings?
Not by itself. Agent actions and search discovery solve different problems. Google says its ordinary SEO best practices remain relevant to generative search features; special AI files are not a prerequisite. See Google's site-owner guidance. Toolkit's plain-text pricing reference makes limits easier to inspect, without promising citations or ranking gains.