DEV Community

Leon
Leon

Posted on

Programs Beat Prompts: AI Forges Deterministic Interface Programs That Run Forever

#productivity

Every time my AI agent automated a website interaction, it was burning tokens to solve the same problem it had already solved last run. Find the API. Locate the selector. Compose the steps. Re-solved, re-paid, every single time.

I built Tap to fix this. The core idea: operating an interface is a solved problem the moment you figure out how. So separate the figuring-out (AI's job, done once) from the executing (a deterministic program's job, done forever).

The Paradigm: Forging 

forge_inspect → forge_verify → forge_save → tap.run
    AI analyzes    AI tests      AI saves     runs forever, zero AI
Enter fullscreen mode Exit fullscreen mode

AI analyzes the live site, creates a .tap.js file, and that file runs forever — no LLM calls, no prompts, no API keys. Runs in <1s. Returns structured data. Same result every time.

Cost model: ~$0.50 in tokens to forge. Then $0.00 at runtime, forever.

The Protocol

Tap defines a minimal, complete contract for interface automation:

8 core operations (irreducible atoms):

eval · pointer · keyboard · nav · wait · screenshot · run · capabilities
Enter fullscreen mode Exit fullscreen mode

17 built-in operations (composed from core):

click · type · fill · hover · scroll · pressKey · select · upload · dialog
fetch · find · cookies · download · waitFor · waitForNetwork · ssrState · storage
Enter fullscreen mode Exit fullscreen mode

8 + 17 = every interaction a human can perform on any interface.

A new runtime implements 8 methods → gets 17 built-ins free. Write a tap once, run it on Chrome, Playwright, macOS native apps — same protocol.

106 Ready-to-Use Skills 

# Install
curl -fsSL https://raw.githubusercontent.com/LeonTing1010/tap/master/install.sh | sh
tap update  # Pulls 106 skills across 50 sites

# Use
tap github trending --limit 5
tap hackernews hot
tap xiaohongshu hot

# Compose with Unix pipes
tap github trending | tap tap/filter --field stars --gt 1000
Enter fullscreen mode Exit fullscreen mode

Skills cover GitHub, Reddit, YouTube, Hacker News, X/Twitter, Medium, arXiv, Bilibili, Zhihu, Xiaohongshu, Weibo, and 40+ more sites.

What a .tap.js looks like 

// API-first: fetch data directly
export default {
  site: "hackernews", name: "hot",
  description: "Hacker News top stories",
  health: { min_rows: 10, non_empty: ["title"] },

  extract: async () => {
    const ids = await fetch('https://hacker-news.firebaseio.com/v0/topstories.json').then(r => r.json())
    const items = await Promise.all(ids.slice(0, 30).map(id =>
      fetch(`https://hacker-news.firebaseio.com/v0/item/${id}.json`).then(r => r.json())
    ))
    return items.map(i => ({ title: i.title, score: String(i.score), url: i.url, author: i.by }))
  }
}
Enter fullscreen mode Exit fullscreen mode

Pure JavaScript. Zero AI at runtime. The health contract means tap doctor can self-verify this tap automatically.

MCP Native

Works with Claude Code, Cursor, Windsurf, and any MCP-compatible agent:

{
  "mcpServers": {
    "tap": { "command": "tap", "args": ["mcp"] }
  }
}
Enter fullscreen mode Exit fullscreen mode

43 tools across 6 categories — run taps, forge new ones, inspect pages, intercept network traffic.

Security

Community taps are untrusted code. Three layers:

  1. Sandbox — Deno Worker, zero permissions (no filesystem, network, env)
  2. Static analysis — 7 CI checks on every PR to tap-skills
  3. Data isolation — secrets and sessions never leave your machine

How It Compares

Tap Browser-Use / Stagehand Playwright
AI at runtime No (forge once) Yes (every step) No (manual scripts)
Detection risk Undetectable Detectable (CDP) Detectable (headless)
Cost model ~$0.50 once, then $0 Tokens per session Free (manual)
Reusable artifacts .tap.js (shareable) None (ephemeral) None
Skills ecosystem 106 across 50 sites None None

GitHub

https://github.com/LeonTing1010/tap

106 skills, 3 runtimes, ~2,000 lines of Deno, zero dependencies.

Top comments (0)