WebMCP Simple Agent

Chrome extension: an in-page chat (plus a side panel) where an AI model executes your prompt using the WebMCP tools the current page exposes — both imperative (document.modelContext.registerTool()) and declarative (<form toolname>). Works on stable Chrome: a MAIN-world content script polyfills document.modelContext, so no origin-trial flag is needed.

Bring your own API key — Anthropic, OpenAI or Kimi, with model selection. Design doc: docs/specs/2026-09-09-webmcp-agent-extension-design.md.

Build & load

pnpm install
pnpm build

1. chrome://extensions → enable Developer mode → Load unpacked → select dist/
2. Extension Options → paste your Anthropic API key (stored in chrome.storage.local)
3. Open any WebMCP page, click the toolbar icon to open the side panel

Manual e2e checklist

Demo page (open demo/index.html via any static server, e.g. python3 -m http.server -d demo):

• Page shows "3 tools registered"; panel shows "3 tools available on this page"
• Prompt "what time is it?" → Claude calls get_time and answers
• Prompt "add a note saying hello, then list all notes" → chained add_note + list_notes, note appears in the page
• Wrong API key → clear error banner pointing at Options

Procmail dashboard (WebMCP toggle enabled in /app/profile):

• Panel lists the dashboard tools on /app pages
• "Crea una broadcast col segmento XYZ, template TPL01, programmala per il 10/10/2026 23:00 UTC" → agent resolves segment/template via list tools, creates the broadcast, and schedule_broadcast opens the dashboard confirmation modal
• Rejecting the modal → agent reports the action was not approved

Development

pnpm test        # vitest
pnpm typecheck   # tsc --noEmit