Updated October 11, 2026. Start by choosing one useful action on your website. A small prototype can be quick; a production implementation also needs permissions, failure handling and browser testing. There is no fixed ten-minute guarantee.
1. Make the ordinary action work
Check the human workflow first. Use named inputs, labels, useful validation and keyboard-accessible controls. A tool should reuse the real application behavior rather than bypassing it.
2. Expose a form or imperative tool
For an existing search form, add a clear tool name and description while retaining its submit handler. For dynamic application logic, use the current document.modelContext interface and an execute callback. The implementation guide provides both examples.
toolautosubmit is optional. Missing it does not mean a form cannot be a tool. Assess automatic submission against the operation's effects and your tested implementation; it never replaces authorization or a needed confirmation.
3. Describe inputs and effects
Use an input schema that represents the actual values, then validate them in the callback and server where applicable. Describe returned untrusted content and consequential actions accurately. Read the current annotation guide before copying fields from MCP.
4. Retain security and publishing controls
Use a secure context and fix mixed content. Keep CSP protections and authorize the scripts you need with nonces, hashes or approved external sources. Do not remove framing restrictions simply to satisfy a detector.
Publish useful descriptions and structured data that match visible content. Optional llms.txt documentation can provide a curated reading path, but is not required for Google indexing. Decide crawler access deliberately rather than automatically allowing every bot.
5. Verify more than a score
Run a free scan to inspect registration and page evidence. The scanner does not execute the target tools. Separately test discovery and invocation in a compatible browser/client, invalid inputs, denied permissions, failures and the original human path. Record the tested browser configuration and ruleset.
A readiness score is a product heuristic, not official certification or search ranking. For actual API availability, follow the browser guide and current draft.