The handshake

Two type changes and a namespace move, inside eight weeks

None of these are large. All of them are the kind of change that produces a tool an agent silently declines to call rather than an error anyone sees. The two marked in pink will reject code that was correct a fortnight earlier.

A spec can churn quietly right up until something reads it

For most of this year the honest thing to say about WebMCP was that nothing consumed it. Nine origin-trial participants were named, deployment was unconfirmed, and every mainstream agent still parsed the DOM. A page that exposed tools was making a statement about where the web is going, not wiring up an integration.

Registering a tool so that all three summers of implementation can read it

Twelve lines. It writes the current shape first, falls back through the namespace move, and only falls back to the stringified schema if the object form was refused, so it never volunteers a deprecated shape to an implementation that wanted the current one.

// Current shape first. Only degrade when the implementation refuses it.
async function registerCompat(tool) {
  const mc = document.modelContext || navigator.modelContext;   // spec moved 21 Jul 2026
  if (!mc?.registerTool) return { ok: false, why: 'no implementation' };

  try {
    await mc.registerTool(tool);                // inputSchema as an object (#241)
    return { ok: true, shape: 'object' };
  } catch (e) {
    // Older implementations typed inputSchema as DOMString.
    await mc.registerTool({ ...tool, inputSchema: JSON.stringify(tool.inputSchema) });
    return { ok: true, shape: 'string', degradedBecause: String(e) };
  }
}
// The thing worth keeping is `degradedBecause`. It is the only place the
// drift is visible, and only if you log it somewhere you will read.

What each row actually does, and what it cannot tell you

Every row attempts a real call and reads the real outcome. None of them infer a capability from a property name, because the presence of registerTool says nothing about which argument shapes it validates. Rows that report never reachable are the honest answer on a browser without WebMCP: no implementation ran, so nothing was learned about it.