The web was built for humans.
Ghost was built for agents.
Ghost turns any website into typed MCP tools your AI agent can call — auto-generated from the live DOM, cached forever, self-healing.
one command · no config · no API keys · 20,000x faster than vision agents
The difference
Structured data vs. screenshot guessing
Screenshot agents send images to LLMs and hope for the best. Ghost returns typed JSON from auto-generated MCP tools.Example: extracting top stories from Hacker News
// Screenshot Agent (GPT-4V, Claude Vision, etc.)
// Take screenshot → Send to LLM → Parse response
const screenshot = await page.screenshot();
// → 1.2MB base64 PNG blob
const response = await llm.chat({
model: "gpt-4-vision",
messages: [{
role: "user",
content: [
{ type: "image", data: screenshot },
{ type: "text", text: "Extract the top stories" }
]
}]
});
// Response: unstructured text, no schema
// Latency: 3-8 seconds per call
// Cost: ~$0.04/screenshot (adds up fast)
// Accuracy: varies by prompt quality// Ghost: Structured MCP tool call
// Auto-generated, typed, cached
const stories = await ghost.execute(
'hackernews_extract_top_stories',
{ limit: 5 }
);
// Returns typed JSON in <1ms (cached):
[
{
title: "Show HN: Ghost – AI browser auto...",
points: 342,
author: "ajgreen",
url: "https://github.com/ajsai47/ghost",
comments: 127
},
// ... 4 more, fully typed
]
// Latency: 0.03ms cached, 27ms first call
// Cost: $0 (runs locally)
// Schema: auto-generated, self-healingScreenshot Agent
Ghost
Under the hood
How Ghost works
Three steps. Any website. Structured data out.
Navigate
Ghost visits any URL and analyzes the DOM structure. It auto-generates typed, callable tools — extract, click, type, scroll — and detects frameworks to adapt its selector strategy.
Execute
A 3-tier execution waterfall finds the fastest path. Cached API replays resolve in under a millisecond. Headless browsers handle fresh pages. The extension handles auth-gated sessions.
Return
Ghost returns typed, structured JSON — not screenshots or blobs. Every field is named, typed, and deterministic. Pipe it directly into your agent's next step.
Interactive
See Ghost in action
Pick a scenario and watch Ghost navigate, analyze, and generate tools in real time.
Benchmarked and proven
Performance
Real numbers from 1,637 tests across 7 benchmark suites.
15ms
Tool generation
552ms
Headless nav avg
325ms
In-page execution avg
3ms
Extension DOM extraction
1,637
Passing tests
100%
Pass rate
Ghost vs alternatives
Use cases
What you can build
From one-off extractions to production monitoring pipelines.
Data Extraction
Extract structured data from any website with a single command.
const results = await ghost.ultra_research({
query: "latest AI papers",
sites: 100,
extract: ["title", "authors", "date", "abstract"]
});
// → 847 papers, JSON + CSV exportNever break again
Self-healing selectors
A 4-strategy pipeline that automatically repairs broken selectors when sites change.
Structural Fingerprint
Each element gets a tag-path > text-hash > landmark > position fingerprint using djb2 hashing.
text 35% · tag 30% · position 20% · landmark 15%
Adaptive Relocation
Multi-factor similarity scoring across tag, text content, attributes, XPath, parent, and sibling context.
Works after major DOM restructuring
Text / ARIA Matching
Falls back to accessible labels, text content, and role attributes. Works even when all CSS classes change.
Framework-agnostic resilience
LLM Healing
Sends DOM context to Claude Sonnet to re-identify the element. Last resort — always succeeds.
AI-powered final fallback
Enterprise-grade infrastructure
Built for scale
Production-ready from day one. Stealth, resilience, and observability baked in.
Anti-Detection & Stealth
Navigate protected sites without getting blocked.
- 9 stealth init scripts on every page
- Cloudflare Turnstile + PerimeterX bypass
- TLS fingerprint impersonation
Proxy Rotation
Rotate IPs automatically to avoid rate limits.
- Round-robin and random strategies
- Auto failure detection & exclusion
- Recovery after backoff period
GHOST_PROXY_URLS
Smart Rate Limiting
Respect site limits automatically.
- 5 concurrent global, 2 per-domain
- Token bucket for API services
- Exponential backoff with jitter on 429/5xx
Tab Pool Management
Efficient browser resource management.
- 10-tab max with LRU eviction
- 5-minute idle auto-cleanup
- Configurable pool size
GHOST_MAX_TABS
Cloud Execution
Run headless in the cloud when local isn't available.
- TinyFish cloud fallback
- SSE streaming for real-time progress
- Auto-failover from local to cloud
Webhook Alerts
Get notified when monitors detect changes.
- HMAC-SHA256 signed payloads
- Retry with exponential backoff
- Per-monitor webhook URLs
Switch
