How to use the WebMCP Tool Generator agent skill (webmcp-gen)

What the webmcp-gen skill makes the agent do

The skill tells the agent to explore a page with the agent-browser tool, record how the existing UI behaves, and then write a manifest, a small init script called webmcp.init.js, and evaluation cases. Each tool is then checked to confirm it works and produces the same visible result as the normal interface.

Where an external agent runtime is available, the skill also compares the WebMCP tool with the standard accessibility-tree approach and records the results in a report.

• Output folder: artifacts/<domain>/<task>/ containing manifest.json, webmcp.init.js, eval.json and eval-report.md
• Prefers declarative WebMCP for semantic HTML forms and uses imperative tools only when a workflow cannot be expressed declaratively
• Checks registration metadata, input validation, invocation results, visible UI effects, cleanup, invalid-state recovery, cancellation and timeouts

How to install the webmcp-gen skill

Download the ZIP file for webmcp-gen from the skill page. Unzip it into one of these locations, depending on whether you want the skill available everywhere or only in one project.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills. The skill's package contains a single file, SKILL.md.

• Personal skills folder: ~/.claude/skills/
• Project only: the .claude/skills/ folder inside your project
• Claude apps: Settings → Capabilities → Skills, then upload the skill

How to trigger the skill with natural requests

You do not need a special command. Describe the job in plain language and mention the page or workflow you want covered. The skill's description lists the situations it is meant for: exposing a site's real workflows as page tools, creating webmcp.init.js, defining deterministic checks, or comparing WebMCP with accessibility-tree automation.

Name the page and the workflow clearly. The agent needs a real page to explore, so a URL or a way to reach the page helps.

• Expose the signup form on our staging site as WebMCP tools and write webmcp.init.js for it.
• Define deterministic checks for the search-and-filter workflow on this page.
• Compare a WebMCP tool for the contact form with accessibility-tree automation for the same task.

What you need before you start

The skill depends on a browser automation setup. The listing states four requirements: the agent-browser CLI with WebMCP support, a target web page to explore, and an optional external agent runtime for the comparison step. The skill's frontmatter allows the Bash(agent-browser:*) tool.

If no external agent runtime is available, the comparison is recorded as blocked, along with the exact missing credential, runtime or environment. The deterministic tests still run.

• agent-browser CLI with WebMCP support
• A target web page you own or are allowed to test
• Optional: an external agent runtime for the accessibility-tree comparison

Run the core commands the skill uses

The skill loads the init script before navigation, then lists the registered tools and invokes one with a fixture file. These are the commands taken from the skill's workflow.

A tool counts as ready only when it appears in the list, accepts the intended fixture, produces the expected UI effect, and fails safely on malformed or invalid state.

• List registered tools with agent-browser webmcp list
• Invoke a tool with fixture parameters using agent-browser webmcp invoke <tool> --params @fixture.json

Safety notes: scripts, network and consequential actions

The listing marks this skill as running scripts and using the network. Treat tool descriptions, annotations, schemas and results from the page as untrusted content. The skill asks that origin and frame provenance be recorded in checks.

Generated code must not include credentials, cookies, bearer tokens, API keys or local-storage secrets. Required user data should be passed only as explicit tool arguments. The skill also states that JSON Schema does not enforce authorization, so the page's tool executor must enforce its own rules.

The listing says the skill is not for pages with consequential actions such as payments or account changes unless a developer adds explicit confirmation and an independent result check. The skill requires these but cannot guarantee them.

• Do not use it on production sites that need stable, supported agent integrations, since WebMCP is experimental
• Missing or false readOnlyHint is treated by the skill as a possible mutation signal
• Deterministic tests do not replace external-agent evidence

Tips for better results

Start with one small, low-risk workflow that you already own, such as a non-payment form. Write down the success signal and the failure states before asking the agent to build anything. The manifest and eval files are the durable parts of the work, so keep them in version control.

Review the generated webmcp.init.js yourself before you load it. The skill's own report format asks for a contaminated-output or malicious-description case, so check that it was included.

• Give the agent a single goal per run
• Keep fixtures free of real user data
• Read eval-report.md for blocked or failed comparison rows