Built for
Anthropic
OpenAI
DeepSeek AI
Cohere
NVIDIA AI
xAI
Anthropic
OpenAI
DeepSeek AI
Cohere
NVIDIA AI
xAI
Anthropic
OpenAI
DeepSeek AI
Cohere
NVIDIA AI
xAI
Anthropic
OpenAI
DeepSeek AI
Cohere
NVIDIA AI
xAI
The difference
Traditional scraping APIs require manual setup for every site. Ghost generates typed tools automatically from the live DOM.Example: scraping product data from Nike.com
// Parse.bot API Configuration
// Sign up → Read docs → Configure selectors
{
"url": "nike.com/products",
"selectors": {
"title": ".product-title",
"price": ".product-price",
"image": ".product-image img",
"link": ".product-card a"
},
"pagination": { "type": "scroll" },
"auth": { "apiKey": "pk_live_..." }
}
// Time: 15 minutes → Failed ✗
// Cost: $50–500/mo + overages// Ghost Auto-Generated Tool
// Install extension → Browse naturally
const product = await ghost.nike_get_product(
"air-jordan-1"
);
// Returns:
{
name: "Air Jordan 1 Mid",
price: "$125.00",
inStock: true,
colorways: 5,
url: "/product/air-jordan-1-mid"
}
// Time: 20ms → Success ✓
// Cached — next call: 0.03ms, 0 tokens
// Cost: $0 (runs locally)Traditional API
Ghost
Switch
Architecture
Three-tier cascade. Each tier is a fallback for the one above. Ghost auto-selects the fastest path for every request.
Recognizes common page layouts on sight. No AI call needed. Covers 80%+ of sites instantly.
If the first approach breaks, it automatically tries five other strategies. Self-healing, no intervention needed.
For truly unusual pages, an LLM refines selectors in the background. Your tools work immediately; accuracy improves async.
Install Ghost and start generating tools in under 10 seconds.
Interactive
Pick a scenario and watch Ghost navigate, analyze, and generate tools in real time.
Intelligence
Ghost uses Claude Opus 4.6 with extended thinking for its most critical operations. Not just API calls — deep reasoning that makes tools smarter over time.
When a website changes its DOM and CSS selectors break, Opus 4.6 reasons about the breakage using extended thinking (chain-of-thought), then generates repaired selectors that adapt to the new layout. Zero human intervention.
Ghost's ultra-research mode uses Opus 4.6 extended thinking to plan multi-site research strategies. The model reasons about which sites to visit, what data to extract, and how to synthesize findings — all with visible chain-of-thought.
Why extended thinking matters: Standard LLM calls generate tools that work 80% of the time. With extended thinking, Opus 4.6 reasons through edge cases — shadow DOM, dynamic class names, framework-specific patterns — producing selectors that survive site redesigns. Ghost's 94.6% success rate is built on this deeper reasoning.
Benchmarks
Every number is from real-world testing against production websites.
Same task. Same site. Wildly different results.
| Parse.bot (YC F'25) | Ghost | |
|---|---|---|
| Nike.com products | 15+ min (failed) | 20ms (0.03ms cached) ✅ |
| Setup | Sign up, API keys, docs | Install extension |
| Cost | $50–500/mo + overages | $0 (local) |
| Maintenance | Manual selector updates | Self-healing |
| Privacy | Cloud (data on their servers) | 100% local |
| API keys | Required | 0 required |
Every other tool runs the same work on repeat. Ghost generates once, caches forever.
| Day 1 | Day 30 | |
|---|---|---|
| Avg generation | 20ms | 0.03ms |
| Tokens used | Minimal | Zero |
| Tools cached | 0 | 1,000+ |
| Maintenance needed | None | Still none |
Parse.bot on day 30? Same as day 1. Same latency. Same cost. Same token burn.
Use Cases
Not “save time.” Not “easy to use.” Specific technical capabilities for specific technical problems.
Ghost-generated tools replace fragile Selenium selectors. Self-healing means tests don't break on redesigns.
Search → filter → select → purchase. Ghost generates a tool for each step. Chain them in your agent.
Generate tools for competitor product pages. Run them on a schedule. Structured pricing data, zero maintenance.
Ghost maps form fields automatically. Fill, submit, and verify across any web application.
No selector updates when sites change. Quality scoring detects degradation. Auto-regeneration on failures.
Extract structured content from documentation sites, wikis, forums. Clean data for your retrieval pipeline.
// Auto-generated Nike.com tool — no config needed
const product = await ghost.nike_get_product("air-jordan-1");
// { name: "Air Jordan 1", price: "$170", inStock: true }
// Competitive monitoring — runs on schedule
const prices = await ghost.amazon_get_price("B09V3KXJPB");
const competitor = await ghost.bestbuy_get_price("6505727");
// Form automation — maps fields automatically
await ghost.salesforce_fill_lead({
name: "Acme Corp",
email: "deals@acme.com",
stage: "Qualified"
});Technical
Works with Claude Code, Cline, Continue, Cursor. Any MCP client gets structured web tools automatically.
data-testid → aria-label → id → semantic HTML → text content. If one breaks, the next takes over.
Tools, sessions, data — everything in ~/.ghost/ on your device. No cloud. No telemetry. No third-party processing.
Stealth mode, auth persistence, multi-tab. Handles anti-bot measures and authenticated sessions.
Instant heuristic generation in 27ms. LLM refines in background. Tools work immediately, improve async.
Quality scoring per selector. Detects degradation automatically. Regenerates broken tools without intervention.