Meta has published developer instructions that let web apps on its Ray-Ban Display glasses act on a wearer’s spoken commands, sparing developers from building their own speech recognition.

The plumbing is WebMCP, a browser interface that installs on document.modelContext before a page’s own scripts run; developers register their tools there.

Meta’s own pitch is compact: “The wearer speaks to Meta AI, Meta AI calls your tools, the app updates. No SDK, no speech recognition, no wake word.”

What Meta shipped is a build-time kit, not a switch a buyer flips. Whether an app qualifies is fixed at launch and holds for the whole session.

Meta stresses the feature is strictly additive: the app has to remain usable by D-pad alone, with no assistant present.

Four coding assistants get plugin support: Muse Code, Claude Code, Codex and Cursor. Meta pushed the manifests to 128.1.0.

Parameters may only be a boolean, integer, number or string; arrays and objects are rejected outright. Meta’s rules also strip schema keywords including enum, minimum and maximum, so they are never enforced. Developers are told to check arguments inside the tool, which must answer an assistant call within ten seconds.

A tool cannot even flag its own failure. Every completed call is logged as success, so an app has to explain the problem inside an ordinary reply. Names may run to 128 characters, though Meta AI re-registers each within a 64-character budget.

Source: Mixed News