WebMCP lets a page hand AI agents a list of callable actions instead of making them guess their way through your buttons from a screenshot. This decides whether you should build any, shapes them if you should, and stops you publishing claims the evidence does not support.
This file is the Claude Code wrapper. The same content is at AGENTS.md in the
repository root for every other agent runtime.
Most sites should not build WebMCP tools yet. Returning that answer with the reasons is a successful run.
Routes
The flow is assess then design then implement then scaffold then verify
then trial. review and claims are usable at any point.
| You want | Route |
|---|---|
| Should we build WebMCP tools at all | assess, read method |
| Scope and shape the tools | design, read method then api-reference |
| Write the actual tool code | implement, read implement |
| Wire it into React, Vue, Angular, Next or vanilla | scaffold, read scaffold |
| Prove it works before shipping | verify, read verify |
| Local flag and origin trial token | trial, read trial |
| Security and correctness review of a proposed tool | review, read api-reference |
| Is this claim safe to publish | claims, read claims-and-maturity |
| Exact API surface, gates, budgets | api, read api-reference |
| Which package, which repo, what licence | read ecosystem |
Working templates are in templates/. Every one is checked by
scripts/check-templates.sh for the deprecated accessor, the annotation contract,
signal threading and name budgets.
Method
- State the maturity position first, not as a closing caveat. Read claims-and-maturity. A reader acts on the recommendation and skims the disclaimer, so the disclaimer cannot carry the load.
- Separate the four questions people conflate: can agents reach this site, can they read it, can they act on it, does any of it affect ranking. WebMCP answers only the third.
- Run the fit assessment. Four tests, in method. Three or four passes is a scoped build. Two is defer with a date. Fewer is no.
- If building, design against api-reference. Enforce the annotation contract and check both silent-failure gates.
- If adjudicating a claim, apply the claim ladder. Return the exact supported wording, and name each failing claim with the reason.
- Write the measurement plan and its limits. No external telemetry exists.
What this will and will not do
It WILL write tool code into your repository, scaffold it into your framework, and tell you how to verify it.
It will NOT:
- Claim WebMCP affects rankings, indexing, crawling, traffic, or conversion. No current evidence supports any of them.
- Deploy. It does not enrol you in an origin trial, set browser flags, install extensions, run a CLI against production, or modify a live site. Those are the owner's actions.
- Present a vendor benchmark as independent validation.
- Invent a statistic, a deployment, a capability, or a currentness claim. Missing
evidence returns
no data. Missing decisions returnneeds_input. - Build before
assesspasses. Most sites should not have tools yet, and returning that answer with reasons is a successful run.
The one fact that governs every marketing claim
Chrome's own documentation:
"Clients and browsers must visit a site directly to know if it has callable tools."
There is no registry, no directory, no crawl path. An agent must already be on your page. WebMCP cannot bring you traffic. It changes what happens to traffic you already have. Third-party directories exist; no agent client is documented as reading them.
Volatility
Every fact here was verified against primary sources on 2026-09-05 and re-verified
2026-09-06 by a three-lane audit that corrected six errors. WebMCP is pre-standard
and moves weekly. consequentialHint was two days old at capture. The Chrome origin
trial ends 2026-11-17.
Re-verify before 2026-12-01, or immediately if any browser ships WebMCP by default. Treat an expired horizon as a blocker, not a warning. Sources and dates are in docs/sources.md.
Evaluation
Synthetic cases in evals: a good fit, a poor fit pushed enthusiastically, a claim adjudication, a security review, and a missing-evidence refusal. Case definitions are not a passing run. Nothing here has been through an independent behavioural evaluation.