WebMCP Gen
Author website-specific WebMCP tools by writing a manifest, compiling it to an init script, and validating that Chrome registers and invokes the tools.
This skill does not call a nested agent. You are responsible for exploring the page, writing manifest.json, and iterating based on validation output.
Setup check
From the skill directory, install dependencies if they are not already installed:
bashcd skills/webmcp-gen pnpm install
This installs the pinned Stagehand package plus the TypeScript toolchain (tsx,
typescript, @types/node) used to run the generated stagehand-example.ts.
Workflow
- Pick an artifact slug with exactly one slash:
text<domain>/<task>
Example:
textexample.com/page-context
- Scaffold the artifact:
bashnode scripts/scaffold.mjs example.com/page-context --url https://example.com
- Explore the target page with the
browseCLI:
bashbrowse open https://example.com --local browse snapshot browse get title browse get url browse get text body browse get html body
Prefer browse snapshot, page text, and DOM inspection over screenshots unless visual layout matters. Use browse stop when exploration is complete.
-
Edit
artifacts/<domain>/<task>/manifest.json. The manifest is the source of truth. -
Compile:
bashnode scripts/compile.mjs artifacts/example.com/page-context
- Generate a runnable Stagehand example (
stagehand-example.ts) and run it withtsx:
bashnode scripts/generate-stagehand-example.mjs artifacts/example.com/page-context npx tsx artifacts/example.com/page-context/stagehand-example.ts
- Validate:
bashnode scripts/validate.mjs artifacts/example.com/page-context
- If validation fails, inspect
eval.jsonandeval-report.md, patchmanifest.json, then compile and validate again.
Manifest contract
json{ "domain": "example.com", "task": "page-context", "url": "https://example.com", "generatedAt": "2026-06-04T00:00:00.000Z", "tools": [ { "name": "example_com_page_context", "description": "Returns page context.", "inputSchema": { "type": "object", "properties": {}, "required": [] }, "implementation": { "kind": "dom", "source": "return { success: true, title: document.title, url: location.href };" }, "fixtureInput": {} } ] }
Authoring rules
implementation.sourceis inserted insideasync (input) => { ... }; write JavaScript statements, not a full function wrapper.- Return a JSON-serializable object.
- WebMCP code runs inside the browser page. Use browser-native APIs:
document,location,navigator, and same-originfetch. - Do not use Playwright, Puppeteer, Stagehand, XPath helpers, or agent/browser commands inside
implementation.source. document.querySelectorandquerySelectorAllmust receive valid browser CSS selectors only.- To find visible text, use
Array.from(document.querySelectorAll(...)).find((el) => (el.textContent || "").includes("...")). - Do not include API keys, bearer tokens, cookies, localStorage secrets, or user credentials.
- Do not use
evalornew Function. - Avoid destructive actions unless the user explicitly asked for them.
- Make implementations defensive: check for missing elements and return structured
{ success: false, error: "..." }responses. - Generated init scripts register WebMCP tools only in the top frame.
Output layout
textartifacts/<domain>/<task>/ manifest.json webmcp.init.js stagehand-example.ts eval.json eval-report.md
To turn the example into a standalone project, scaffold a Stagehand app with
npx create-browser-app and drop the generated webmcp.init.js into it (load it
with page.addInitScript({ path: "webmcp.init.js" })).

