Skill-based agent orchestrator with a web UI. Takes a natural-language prompt, routes it to the right skill category via LLM, and spawns parallel scripts to fetch results from multiple providers.

server.ts                         HTTP + WebSocket server
router.ts                         LLM-powered prompt → category routing (Claude Haiku)
orchestrator.ts                   Discovers skills, extracts scripts, runs in parallel
web/
  index.html                      Home screen with category cards
  app.ts                          Frontend: WebSocket client, renderers, filters
  renderers/
    flights.ts                    Flight card renderer + flex dates
    trends.ts                     Trend card renderer + grouped view
    default.ts                    Fallback JSON renderer
skills/
  {category}/
    config.json                   { maxConcurrent, args, requiredFields, routerHints, renderer }
    {provider}/SKILL.md           Instructions + embedded script (## Script block)

The orchestrator discovers skills/{category}/config.json, extracts ## Script code blocks from each SKILL.md into runnable .mjs files, spawns one Node process per skill (respecting maxConcurrent), and streams JSON results via WebSocket.

npm install
npm run dev

Open http://localhost:3000 — pick a category (Flights or Trends), type a prompt or click an example.

Set ANTHROPIC_OAUTH_REFRESH_TOKEN or place .auth.json in the root.

Trends supports two modes:

• Global: omit query to get what's trending right now across all platforms
• Query-based: provide a topic (e.g., "AI trends", "teenager trends") to search each platform

1. Create skills/{category}/{provider}/SKILL.md with frontmatter:

   ---
   name: provider-name
   description: What this provider does
   ---

2. Add a ## Script section with a ```javascript code block
3. Script must output JSON to stdout: { "site": "Name", "url": "...", "{items_key}": [...], "partial": false }
4. Args are passed as CLI flags: --from EZE --to JFK --departure 2026-04-15

1. Create skills/{category}/config.json with:
   • maxConcurrent: parallel limit
   • description: what this category does
   • args: JSON schema for arguments
   • requiredFields: which output fields to validate
   • routerHints: guidance for the LLM router
   • renderer: which frontend renderer to use (or falls back to default)
   • items: key name for the results array
2. Create provider directories with SKILL.md files
3. Optionally create web/renderers/{category}.ts and register in web/app.ts
4. The router and orchestrator auto-discover new categories

# Run via CLI (no web UI)
npx tsx agent.ts "flights from Buenos Aires to New York next month"

# With model override
MODEL=claude-opus-4-6 npx tsx agent.ts "what's trending right now"

npm install
npm run dev          # Dev mode with hot reload
npm run build:web    # Build frontend only
npm start            # Production mode