/webmcp-kit:implement

Reference for /webmcp-kit:implement — add WebMCP tools to a website codebase, with a reviewed plan before any code.

Adds WebMCP tools to the website repo you’re working in. Plans first, writes only after you approve, verifies in a browser.

/webmcp-kit:implement

Run it from the root of the site’s repo. No arguments.

When to use it

• Making a site agent-ready for the first time.
• Re-running after a redesign or new feature — existing tools keep their identity via stableKey.
• Someone asked you to “add WebMCP tools” or “expose this site to browser agents”.

What you review

The plan (Phase D — see How it works for all six phases) shows, per tool:

• Name and description, verbatim. The description is what an agent reads to decide when to call the tool — it ships exactly as approved, so word it like it matters.
• Wiring — the concrete code path the tool will call (a data-layer function, a same-origin route).
• Availability — which pages and auth states register the tool.

Tools come in two confidence levels:

• Decided — a must-have journey with a clearly safe wiring path. Listed for your review.
• Needs your input — anything ambiguous. Comes as a specific question with a stated default, never a silent guess.

Journeys with no safe client path are listed as needs developer wiring — the plugin tells you what’s missing instead of faking it.

What it writes

After approval, in your repo’s own language and conventions:

• Tool modules — one or more files of defineTool definitions. Side-effect-free: importing them registers nothing.
• One entry module per scope — the single place that calls registerTools and owns the tools’ lifetime.
• The SDK dependency — wired the way your repo needs (registry package or vendored copy; transpilePackages added for Next.js when required).

Then verification runs (same ladder as the verify skill), and the work lands on a webmcp/tools-v0 branch with the plan and per-tool results as the PR body.

Headless runs

With no human available to approve (CI, non-interactive sessions), the skill proceeds on stated defaults and records everything: the plan verbatim in the PR body, plus each assumption — the question it would have asked, the default it took, and why.

Interactive loop

If you prefer reviewing in a live browser page instead of chat, say so — the skill renders the plan in an Explorer page where approval is a click. State lives in a git-tracked .webmcp/ folder.