Add Webmcp

Add WebMCP

Turn capabilities already implemented by a web app into maintained, first-party WebMCP tools. Modify the target codebase and its tests; do not introduce a hosted proxy or third-party runtime.

Compatibility: the bundled Stagehand validator requires Node.js 22.18 or newer. Validation needs Chrome/Chromium locally or BROWSERBASE_API_KEY for a publicly reachable preview.

Be verbose as you work: report what each step found as you go, not only in the final report.

Use webmcp-gen instead when the requested output is a standalone init script derived from a live URL. This skill starts from source code and integrates tools into the application.

1. Establish the application boundary

Read the target repository's instructions, package manifests, framework configuration, and current git status. Preserve unrelated changes.

Set ADD_WEBMCP_SKILL_DIR to the directory containing this file and run the bounded scanner:

bashCopy code
node "$ADD_WEBMCP_SKILL_DIR/scripts/scan-codebase.mjs" "$TARGET_REPO"

Treat scanner results as leads, not conclusions. In a monorepo, identify the actual browser app and the server packages it calls before editing.

2. Build a capability inventory

Trace each candidate from its user-visible entry point through the client handler, validation schema, server boundary, authorization checks, side effect, and returned state. Look at:

• routes and screens;
• forms and their submit handlers;
• server actions, API handlers, RPC procedures, and service clients;
• Zod, Valibot, Yup, Joi, JSON Schema, or equivalent validators;
• authentication, authorization, CSRF, idempotency, rate limits, and audit hooks.

Prefer complete user tasks such as search_catalog or save_draft, not a mechanical tool per endpoint. Exclude internal/admin-only operations, authentication bypasses, raw database access, secret-bearing operations, and capabilities the UI does not grant the current user.

For each selected tool, record its source files, existing validation and authorization boundary, side effects, risk class, confirmation behavior, and a safe fixture input. Read references/implementation-and-validation.md for the detailed inventory and framework patterns.

3. Design the tool contract

• Use a stable verb-noun name and describe the user-visible effect, prerequisites, and important exclusions.
• Derive JSON Schema from the application's existing validator or domain type. Do not invent a second, looser contract. Close object schemas with additionalProperties: false and make the execute-time parser reject unknown fields too; a closed discovery schema backed by a permissive runtime parser is not a closed contract.
• Return compact JSON-serializable domain results. Do not return DOM nodes, credentials, cookies, tokens, or entire HTML documents.
• Call the same client/service boundary as the UI so existing validation, authorization, observability, and business rules remain authoritative.
• Validate again inside the handler. Agent-provided input is untrusted.
• Never echo the request back as the result. Read the outcome from the application's own state, and where that state is updated asynchronously (React and most reactive stores do not reflect a change on the next line), poll until it settles before reading, then report whatever is actually true. Echoing turns a silent no-op into a passing test.

Assign annotations deliberately:

Registration uses the current WebMCP hint names readOnlyHint and untrustedContentHint. Stagehand v4 normalizes discovered annotations to readOnly, untrustedContent, and, for declarative forms, autosubmit. These are hints to the browser or agent, not security enforcement. The application must enforce permissions, validation, confirmation, idempotency, and replay protection.

4. Integrate with the application

Use the runtime model context exposed by the browser:

jsCopy code
const modelContext = navigator.modelContext || document.modelContext;

Keep both accessors: current Chrome exposes document.modelContext as a native ModelContext while navigator.modelContext is undefined, so the fallback is load-bearing rather than defensive. The surface is browser-provided and present on any page, so the application ships no polyfill.

Register imperative tools from a client-only root/provider after the application is ready. registerTool returns a promise and is idempotent by name — re-registering replaces the previous definition rather than duplicating it, and there is no unregister handle — so remounting and hot reload are safe without teardown. Use declarative form attributes when an existing form already represents the exact task and preserving a visible review step is valuable.

Do not duplicate server business logic in the tool executor. Do not weaken CSRF, same-origin, auth, or confirmation checks to make a smoke test pass. Never embed secrets in browser code.

5. Verify the implementation

Run the target's focused tests, typecheck, and production build. Then create a small webmcp.e2e.json with every expected tool. Discovery is mandatory; invocation is opt-in per test case and must use synthetic or disposable data.

Install the validator dependencies once:

bashCopy code
pnpm --dir "$ADD_WEBMCP_SKILL_DIR" install --frozen-lockfile

Validate localhost with a Stagehand-launched local browser:

bashCopy code
node "$ADD_WEBMCP_SKILL_DIR/scripts/validate-stagehand.mjs" \
  --url http://localhost:3000 \
  --config "$TARGET_REPO/webmcp.e2e.json" \
  --local

Local runs are headed by default so the browser is visible while it validates; pass --headless for CI or unattended runs.

If discovery reports zero tools, check the host before suspecting the code: dev servers commonly bind localhost only, so --url http://127.0.0.1:PORT finds nothing while http://localhost:PORT works. The failure looks identical to tools never registering.

Use --browserbase only for a publicly reachable deployed preview. The validator uses Stagehand v4's real page.tools(), tool.invoke(), and invocation.result() path. It refuses consequential invocations unless --allow-consequential is explicitly supplied.

An injected init script is useful for testing the validator itself, but it is not proof that the target app ships its own tools. Final application proof must run without --init-script.

6. Adversarially verify the tools actually worked

Step 5 proves each tool is discoverable and that its executor ran. It does not prove the tool did what it claimed, and a passing config is not evidence of a sound contract. Run this step last, after step 5 is green, and drive it yourself against the live page rather than encoding it in webmcp.e2e.json — the point is to probe inputs the author did not anticipate.

Drive the page with a persistent browser session so probes accumulate against real state. The browse CLI is the lightest option — one global install, and the session survives between commands:

bashCopy code
browse open http://localhost:3000 --session probe --local --headed
browse eval --session probe '(async()=>{const mc=document.modelContext;const t=(await mc.getTools()).find(x=>x.name==="my_tool");try{return "ACCEPTED "+JSON.stringify(await mc.executeTool(t,JSON.stringify({/* probe input */})));}catch(e){return "REJECTED";}})()'
browse screenshot --session probe --path /tmp/effect.png

Three things will cost time otherwise: browse open defaults to a cloud browser that cannot reach localhost, so --local is required; browse eval reliably accepts only single-line scripts, so run one probe per command; and executeTool takes the tool object plus arguments as a JSON string — a plain object fails with "Failed to parse input arguments".

Assert on rejected-versus-accepted, not on error text. The WebMCP layer replaces a handler's message with a generic invocation-failed string, so a precise reason never reaches the caller.

Use the discovered tool.inputSchema as the thing under test, not as the source of truth. For each tool, probe:

• Schema closure. Invoke with an extra field the schema does not declare. A tool honoring §3 rejects it. Acceptance means the closed contract is decorative.
• Required fields. Omit a required property. The invocation must fail; a Completed status carrying a null or partial result is worse than an error, because the agent believes it succeeded.
• Types and constraints. Send a string where the schema says number, an out-of-range value against minimum/maximum, and a value outside an enum. Silent coercion or echo-back means the handler never validated.
• Error honesty. Confirm a tool that should fail reports a non-Completed status rather than returning a success-shaped body.
• Annotation honesty. Compare each tool's real annotations against the risk you assigned in step 3. A pure lookup advertising readOnly: false, or a mutating tool advertising readOnly: true, is a defect even though discovery passes.
• Clean rejection. After the malformed probes above, re-read the application's state. A rejected call must leave nothing behind; partial state from a half-applied invocation is a defect the accept/reject result alone will not surface.
• Consequential gating. Never invoke these. Verify the tool declares its risk, that no declarative toolautosubmit is present, and that the app's own confirmation control still stands between the agent and the effect.

Then verify the effect, not the return value. Invoke the tool, then inspect the application independently — DOM assertions for rendered state, a screenshot when the surface is a canvas or chart. A handler that returns {saved: true} without changing anything passes step 5 and fails here. When the surface has no readable DOM, have the tool read back from the application's real store so the returned value is grounded in actual state rather than composed by the executor.

Treat every discrepancy as a defect in the application or the tool contract, and fix it there. Do not loosen a schema, downgrade an annotation, or delete a probe to make this step pass.

7. Report the result

List the capabilities considered and explain exclusions. For each implemented tool, report its contract, backing code path, risk/confirmation treatment, actual Stagehand discovery/invocation result, and the step 6 adversarial probes it survived. State any environment or browser support not tested.

For a comparative benchmark, quality audit, or scored evaluation, read references/quality-rubric.md. Apply its qualification gates before reporting numerical scores; do not let a high diagnostic score hide fabricated behavior, an unsafe consequence boundary, or missing production discovery.