webmcp-agents

Use this bundle as the entrypoint for WebMCP work. It keeps one page-local operating model while routing specialized work to the smallest relevant skill under skills/.

Route the request

• webmcp-agent-browser — inspect, discover, execute, refresh, and verify like an agent browser.
• webmcp-core — normative registration, discovery, execution, lifecycle, and cancellation.
• webmcp-tool-design — tool schemas, UI↔UX journeys, state, confirmation, and recovery.
• webmcp-declarative — exploratory HTML-form tools and preview behavior.
• webmcp-frameworks — React, Angular, TypeScript, SSR, and lifecycle integration.
• webmcp-design-md — DESIGN.md tokens, Lexer → AST → semantic model → Renderer workflows.
• webmcp-security — origin, permissions, privacy, untrusted content, and consequential actions.
• webmcp-evals — selection, chaining, outputs, reliability, and prompt-injection evaluation.
• webmcp-service-workers — exploratory background routing and multi-tab architecture.
• webmcp-setup — browser flags, origin trials, support checks, and fallback behavior.
• webmcp-maintainer — source authority, validation, publication, and stale-content checks.

Shared operating loop

intent → observe live page → discover semantic tools → validate input
→ execute page-owned action → observe fresh state → verify result or recover

Prefer WebMCP tools when they cover the user goal. Keep DOM observation and direct browser actuation as explicit fallback and recovery primitives. Human-visible UI, accessibility, loading, errors, confirmation, cancellation, and agent results must remain one coherent journey.

Authority

Use index.bs for normative behavior. Treat the declarative API, service-worker material, and external DESIGN.md format as exploratory or version-gated unless the source says otherwise.