webmcp-eval

For consultants, agencies, and operators preparing client sites for agentic traffic

Know whether a client page is ready for agents

Get an actionable, detailed review and build plan for making any public page ready for AI agents, based on your content, not generalized advice. Point the eval at one public URL and you see the tools the page already registers and whether to keep, rename or fix them, the tools to add (each tied to a link or form on the page), and what must never be exposed (checkout, place-order, anything irreversible).

First evaluation free — no card. On the connector consent page, choose Try one free evaluation. One per person. The $26 starter below is the paid pack.

Free scanners give you a score. This gives your agent the build list — so the next step in the same conversation is the registration code, not a meeting about what the score means. Built for people who sit between a client website and the agents that will soon arrive uninvited.

  1. Name the page. A services page, an audit offer, a product URL — one address.
  2. Read the verdict. Capability, page intent, tools observed, tools recommended, do-not-expose list.
  3. Brief the client. What to publish in WebMCP, what to keep human-only, what is already fine.

One result, wherever you work

Results your agent can act on

Every evaluation comes back with the same structured results. Each recommended tool carries its input schema, a relation (keep, rename, fix_description, new, needs_owner_input) and the evidence it rests on: the page, the link or form, and the HTTP status. The result also says what was checked and what was not.

In Claude or ChatGPT

Ask for the evaluation, then keep going: “write the WebMCP registration code for the ready offers” or “brief the client in plain English”. The model works from the plan, not from a guess at the DOM.

In your own app

Store results per client page, compare runs before and after a build, or feed your own reports and dashboards. Review the sample for the fields and data provided to your favorite model or custom app.

From a CLI or CI

Any MCP client with a Bearer usr_live_… header can call it — Claude Code, Cursor, or a script. Re-check client pages on a schedule and flag a build when a registered tool comes back fix_description or needs_owner_input. Setup.

Worked example

Say this in Claude.ai or ChatGPT

Evaluate https://www.sunbusinessgroup.com/ai/ai-audit.html with evaluate_webmcp.
Return page_intent, recommended_offers, checked, not_checked, and do_not_expose.

What you should see on that URL:

$2 when that call is billed: the page loads and the analysis runs. The agent returns structured results — see a full sample result and the human brief (print to PDF). Connect Claude, ChatGPT, or Grok.

Who this is for

  • Consultants standing up WebMCP on a client property
  • Agencies that need a repeatable readiness pass before a build
  • Operators who want agents to book and answer — not wander the DOM

Not billed if the page doesn't load (no 2xx response) or neither page rendering nor model analysis succeeds. You still get the result, marked billed: false, and the $2 goes back to your balance.

Custom eval kits are available for larger volumes.

First evaluation free

First evaluation free — no card. One per person. On a connector consent page, choose Try one free evaluation.

ChatGPT Business, Cursor, and other hosts without a consent page: email [email protected] for a one-time pass link.

Connect Claude.ai, ChatGPT, or Grok

Same URL everywhere: https://mcp.webmcp-eval.com/mcp. Claim one free evaluation, or buy a key, then add the connector and paste the example prompt above. Every /mcp call needs a key or an OAuth token. With a key, get_credits is not billed. Only a billed evaluate_webmcp is $2. Results that come back billed: false cost nothing.

Full steps, Cursor snippet, and 401 checklist.

Pages we fetch, and their public HTML, may go to our subprocessor (Firecrawl) so JavaScript-rendered sites can be read. Do not send secrets or private URLs. Terms. Privacy.