What it is
ScrapeGraphAI replaces CSS selectors with a language model. You give it a URL (or a local HTML, XML, JSON or Markdown file) and a plain-English prompt, and a pipeline of steps, which it calls a graph, fetches the page with Playwright and asks an LLM to return the fields as JSON. Ready-made graphs cover one page, many pages, the top search results and generating a reusable Python scraper.
It works with hosted models from OpenAI, Groq, Azure, Gemini, MiniMax and others, or with local models through Ollama, and integrates with LangChain, LlamaIndex, CrewAI, Agno and low-code tools. An MCP server is available.
The library is MIT-licensed and sends telemetry by default; in the current code that includes the prompt, schema, page content, model response and URL of each run. The same company sells a hosted API with proxies and anti-bot handling, from a free tier to $500 a month.
Our review
ScrapeGraphAI solves extraction, not crawling, and its default telemetry is the first thing a team should change. Reading telemetry.py in v2.3.1, each successful run with a schema posts the user prompt, the JSON schema, the scraped page content, the model's response and the URL to a tracing endpoint hosted on Render. The README describes this as anonymous usage metrics; it is closer to a copy of the work. Setting SCRAPEGRAPHAI_TELEMETRY_ENABLED=false stops it.
The extraction itself is the draw. Prompt-driven graphs survive site redesigns that break selector-based scrapers, and swapping a hosted model for a local Ollama model is a one-line config change. The project has about 32,200 GitHub stars, eight open issues and a release roughly every two to three weeks, most recently v2.3.1 on 6 October 2026.
Cost scales with pages. In the library every page is a model call; on the hosted API an Extract request costs 5 credits, so the $100 Growth plan's 100,000 credits buys about 20,000 extractions, or half a cent each. Because a language model reads each page, it can invent or reshape a field that was never there, so outputs need schema validation.
The open-source library has no proxy network or anti-bot layer; protected sites fail unless you bring proxies, and the hosted API reserves proxy rotation for Growth and above. For turning whole sites into Markdown for a RAG index, Crawl4AI or Firecrawl is faster and cheaper, because they do not need a model to read every page.
Where it earns its place
A data engineer at a price-comparison startup
Pull price, stock and spec fields from 300 retailer pages that change layout often
A SmartScraperGraph per page with a Pydantic schema replaces brittle selectors. Each page costs one model call, so a small hosted model or local Ollama keeps the bill down, and telemetry should be disabled first.
A sales-operations analyst at a 50-person B2B firm
Build a contact list from a few hundred company websites
The hosted API's Extract endpoint at 5 credits a page fits the $20 Starter plan's 10,000 credits for about 2,000 pages a month, with no infrastructure to run.
A developer on a regulated team
Extract fields from internal HTML reports without data leaving the network
The library with a local Ollama model and SCRAPEGRAPHAI_TELEMETRY_ENABLED=false keeps everything on-premise; without that flag, page content goes to the maintainers' endpoint.
Before you commit
- Setup
- pip install scrapegraphai, then playwright install for the headless browser; a first extraction is a dozen lines of Python with an API key or an Ollama model pulled locally.
- Integrations
- OpenAI, Groq, Azure, Gemini, MiniMax and Ollama models; LangChain, LlamaIndex, CrewAI and Agno; Zapier, Pipedream and Bubble for the hosted API; an MCP server.
- Data and privacy
- MIT licence. Telemetry is on by default and, per the source code, sends the prompt, schema, page content, LLM response and URL; disable it with SCRAPEGRAPHAI_TELEMETRY_ENABLED=false. No GitHub security advisories published. The hosted API claims SOC 2 Type II; its retention terms are not on the pricing page.
- Cost at scale
- Self-hosted, the cost is model tokens per page plus your own proxies. Hosted: Starter $20 for 10,000 credits, Growth $100 for 100,000, Pro $500 for 750,000; Extract is 5 credits a page and stealth mode adds 5.
- Limits
- Hosted rate limits run from 10 requests a minute on Free to 5,000 on Pro. The library lacks crawling at scale, proxy rotation and anti-bot handling, which Firecrawl and Crawl4AI cover better.
What's good
- Prompt-based extraction that survives layout changes
- Runs fully local with Ollama models
- MIT licence and frequent releases
- Hosted API with a clear per-request credit price
What's not
- Default telemetry sends prompts, page content and responses
- A model call on every page
- Can hallucinate fields without schema validation
- No proxies or anti-bot in the library
Best for
Developers extracting typed fields from a modest number of varied web pages.
Alternatives
ScrapeGraphAI: common questions
- Is ScrapeGraphAI worth it?
- Good at pulling a few typed fields from messy pages, but it ships your scraped content to its makers by default and pays a model call for every page. We rate it 3.6 out of 5.
- How much does ScrapeGraphAI cost?
- Library free under MIT; you pay for LLM tokens. Hosted API: 500 one-off free credits, then $20, $100 or $500 a month; Enterprise custom.
- What are the downsides of ScrapeGraphAI?
- Default telemetry sends prompts, page content and responses; a model call on every page; can hallucinate fields without schema validation; no proxies or anti-bot in the library.
- Who is ScrapeGraphAI best for?
- Developers extracting typed fields from a modest number of varied web pages.
- How much does ScrapeGraphAI cost for a team or in production?
- Self-hosted, the cost is model tokens per page plus your own proxies. Hosted: Starter $20 for 10,000 credits, Growth $100 for 100,000, Pro $500 for 750,000; Extract is 5 credits a page and stealth mode adds 5.
- Is ScrapeGraphAI safe for company data?
- MIT licence. Telemetry is on by default and, per the source code, sends the prompt, schema, page content, LLM response and URL; disable it with SCRAPEGRAPHAI_TELEMETRY_ENABLED=false. No GitHub security advisories published. The hosted API claims SOC 2 Type II; its retention terms are not on the pricing page.
- What are the best ScrapeGraphAI alternatives?
- The closest alternatives we have reviewed are Crawl4AI and Firecrawl.
Answers checked Oct 23, 2026.