Darko893/haunt-mcp-server
📇 ☁️ - Haunt web extraction for AI agents: structured JSON or clean Markdown from permitted public pages, including Cloudflare-protected ones, with honest blocked/login/captcha errors instead of fabricated data
Quick Install
{
"mcpServers": {
"darko893-haunt-mcp-server": {
"command": "npx",
"args": [
"-y",
"darko893-haunt-mcp-server"
]
}
}
}Using an AI coding agent (Claude Code, Cursor, etc.)? Copy a ready-made prompt that tells it to fetch the setup instructions and install this server for you.
Documentation Overview
Haunt MCP Server
Web extraction for AI agents. Give Claude, Cursor, Windsurf, or any MCP client a clean tool that turns any permitted public web page into structured JSON or Markdown from a plain-language prompt, and returns an honest machine-readable failure instead of fabricated data when a page is blocked.
Website · Docs · Get a free key · Python SDK
Built for agent workflows that need product data, competitor prices, article text, page metadata, lead lists, or research snippets, without brittle CSS selectors. Try it with no API key using the try_demo_extract tool. Free tier: 1,000 credits a month, no card.
Quick Start
1. Add Haunt to your MCP client
This is the block to copy.
{
"mcpServers": {
"haunt": {
"command": "npx",
"args": ["-y", "@hauntapi/mcp-server"],
"env": {
"HAUNT_API_KEY": "your-api-key"
}
}
}
}
Get a free API key, no card needed: https://hauntapi.com/?utm_source=github&utm_medium=readme&utm_campaign=mcp_server#signup
Prefer a CLI? npx -y --package @hauntapi/cli@latest haunt-cli init prints the
same block for Claude, Cursor, Windsurf and other MCP clients. It prints config,
it does not edit your files, and it is not itself the server.
2. Prove it is wired in, no key needed
Call try_demo_extract first. It returns a fixed sample extraction JSON and trace
plus docs, signup, pricing and free-tier links, without using credits.
Use Haunt's try_demo_extract tool and show me the signup and docs links.
3. Ask for real data
Use Haunt to extract the product name, price, availability, and review count from https://example.com/product
For Markdown output:
Use Haunt extract_markdown to turn https://example.com/docs into clean Markdown.
Reddit, with no Reddit app and no OAuth:
Use Haunt read_reddit to get the top 10 posts from r/webscraping.
Capability boundaries
Haunt renders pages in a real browser and reads many Cloudflare-protected and bot-walled pages. It does not go past login walls, paywalls, or CAPTCHA human-verification pages, and it does not promise universal extraction. Those return a typed failure signal (login_required, captcha_required, not_found, upstream_fetch_failed) instead of fabricated data, and a failed read is not charged.
Tools
try_demo_extract
No-key activation check. Returns a fixed sample extraction JSON/trace, Haunt's demo endpoint, docs, signup, pricing, MCP info route, and free-tier details. Use this first when a user has installed the MCP server but has not added HAUNT_API_KEY yet.
extract / extract_url
General-purpose extraction from permitted public web pages. extract and extract_url are aliases so hosted MCP docs and local stdio package users can follow the same wording.
Use it for:
- Product names, prices, stock status, reviews
- Competitor pricing pages
- Directories and lead lists
- Job boards
- Research pages
- Supported permitted public pages where you want clean JSON instead of HTML
It accepts optional response_format values: json (default), markdown / md, and raw_html / html.
extract_markdown
Return clean Markdown from a permitted public page. Use this when the agent or workflow wants readable page text for RAG, notes, docs ingestion, or saving as a .md file instead of structured fields.
extract_article
Extract article fields from news, blog, and editorial pages.
Returns title, body text, author, and publish date when available.
extract_metadata
Extract page metadata including title, description, Open Graph tags, Twitter Card tags, canonical URL, and related metadata.
read_reddit
Read Reddit posts or a post's comment thread as structured JSON. No Reddit app, no OAuth, no developer account. Accepts a subreddit URL, a post URL, or shorthand like r/webdev. Set comments to true to read a thread.
Costs 1 credit, and a blocked read costs nothing.
company_profile
Turn a company website URL into a structured profile: what the company does, its product, and the contact and social details published on the site. Read live from the site rather than a stale third-party database.
Costs no credits.
github_repo
Return normalised metadata for a public GitHub repository: description, stars, forks, language, licence, topics, and last push.
Costs no credits.
get_usage
Check current plan, monthly credit limit, used credits, reserved credits, and remaining credits. Use this after a live extraction to see what was charged.
Why Haunt
- Reddit posts and comments with no Reddit app, OAuth, or developer account
- Natural-language prompts instead of fragile CSS selectors
- Supported fetch paths for JavaScript-heavy pages
- Challenge-aware extraction with machine-readable verification signals (
error_code,captcha_provider,requires_human_verification) - Clean JSON output for agents, databases, and workflows
- Clean Markdown output when the job is page text for agents, RAG, notes, or
.mdfiles - Free tier for testing
Pricing
| Plan | Credits | Price |
|---|---|---|
| Free | 1,000/mo | £0 |
| Starter | 10,000/mo | £19/mo |
| Pro | 30,000/mo | £49/mo |
| Scale | 80,000/mo | £99/mo |
Credits are not one-to-one requests. Simple public/non-LLM output usually uses 1 credit, normal structured extraction 2, browser-rendered or authenticated extraction 4, and heavy/screenshot extraction 8. Failed, blocked, login/CAPTCHA, provider, and server failures do not burn credits.
Upgrade: https://hauntapi.com/#pricing