Answers · AEO & AI Search

What is WebMCP and what does it let an AI agent do on a website?

Last reviewed: · Strategic AI Architects Data Desk · reviewed by Mike Moore

WebMCP is a website exposing callable tools to AI agents — so that an agent can do something on the site (run a calculator, look up live data, request an audit) instead of scraping its prose — and it lets an agent discover a site's tools, read each tool's input schema, and invoke them with structured arguments over the Model Context Protocol (MCP). It exists today in two forms: a proposed browser API, and site-served HTTP endpoints, which is the form running on our production sites right now.

The qualifier is that "WebMCP" currently names an emerging landscape, not a finished standard. The browser-API form — navigator.modelContext, co-authored by Google and Microsoft engineers in the W3C Web Machine Learning Community Group — is published as a Draft Community Group Report (latest draft August 17, 2026) and is explicitly "not a W3C Standard nor is it on the W3C Standards Track"; Chrome previewed it in February 2026 and is running an origin trial. The second form needs no browser at all: the site serves a discovery manifest at /.well-known/mcp.json and an MCP endpoint at /mcp that answers JSON-RPC tools/list and tools/call — callable today by any AI agent that can make an HTTP request.

What an agent can concretely do — live, today

These are production deployments, and every row was verified by calling tools/list on the live endpoint on August 18, 2026:

Live WebMCP endpoints on SAA-built production sites (verified 2026-08-18)
SiteEndpointWhat an agent can call
bigsiouxlife.com/mcpA real life-insurance coverage calculator (calculate_life_insurance_need: income, years, mortgage, debts → an educational coverage estimate), plus coverage-type explainers, service-area lookup, and page search.
ichrasavings.com/mcpICHRA savings tools on plan-year-2026 data: ichra_county_savings (county vs. small-group benchmark premiums), ichra_state_summary, and ichra_employer_estimate — the same math as the on-site qualifier funnel, returned as an illustrative range, never a quote.
bigsiouxbenefits.com/mcpA Medicare Advantage plan finder (medicare_plan_finder: plans with drug coverage by ZIP — plan count, lowest premium, top-rated plan).
strategicaiarchitects.com/mcpThis site: aeo_audit_request (request a free AEO audit for a URL), a HIPAA tracking-scan request, and a strategy-call booking-link tool.

The mechanism in one exchange: an agent POSTs {"method":"tools/list"} and receives every tool with its JSON input schema; it then POSTs tools/call with a tool name and arguments and receives structured data back. No scraping, no guessing at form fields. Discovery starts from llms.txt, whose "Tools (agent-callable)" section points at the manifest.

Why we ship it: pillars one through three of AI-citation work make a site readable by engines; a callable tool makes it usable by an agent — and in the insurance vertical, agent-callable web tools are still essentially absent, which is precisely why being the retrievable, verifiable source is a first-mover position rather than a checkbox.

Sources

  1. W3C Web Machine Learning Community Group. "WebMCP" — Draft Community Group Report, navigator.modelContext, editors from Google and Microsoft; "not a W3C Standard nor is it on the W3C Standards Track." Latest draft August 17, 2026, verified live 2026-08-18. webmachinelearning.github.io/webmcp.
  2. Model Context Protocol. Specification — the tools/list and tools/call primitives the HTTP form speaks. Verified live 2026-08-18. modelcontextprotocol.io.
  3. Strategic AI Architects. Live endpoint verification, 2026-08-18: JSON-RPC tools/list calls against bigsiouxlife.com/mcp, ichrasavings.com/mcp, bigsiouxbenefits.com/mcp, and strategicaiarchitects.com/mcp, each returning the tool inventories described above.

Related questions

Is WebMCP an official web standard?

Not yet. The browser-API form is a Draft Community Group Report in the W3C Web Machine Learning Community Group — its own boilerplate states it is not a W3C Standard and not on the W3C standards track. Chrome shipped an early preview in February 2026 and it is running as a public origin trial; Microsoft co-authored the spec, and other engines are engaged without committed timelines. The site-endpoint form we run needs no browser support at all — it's ordinary HTTP speaking the MCP protocol, callable today by any agent that can make a POST request.

What's the difference between WebMCP and just having an API?

Discovery and a shared protocol. A conventional API requires the caller to have read your docs in advance. MCP gives agents one uniform handshake — tools/list returns every tool with a name, description, and JSON schema for its inputs, and tools/call invokes one — so an agent that has never seen your site can discover what it can do and do it, the same way on every MCP site. The /.well-known/mcp.json manifest is the doorbell that tells a visiting agent the endpoint exists.

Is it safe to let AI agents call tools on an insurance website?

It is if you scope it the way we do: every tool at our web edge is read-only or estimate-only — search pages, explain coverage types, compute an educational estimate, look up county savings data. No writes, no PII storage, no quoting that binds anyone; outputs carry 'verify with a licensed agent' framing because this is YMYL territory. Lead capture stays behind the human path. An agent can learn and estimate; it cannot transact on someone's behalf without a person in the loop.

Want to know where your own site stands? Run the free Audit — a live AEO audit plus a HIPAA tracking scan — or browse all answers.