Today we are shipping CrawlForge v4.2.2, our biggest release since launch. It brings three new tools, a standalone command-line interface, and a quiet shift in how we think about web scraping for AI: most of it should run locally, on your own machine, without API keys.
This post is the umbrella for everything in 4.2.2. Three deep-dive guides follow in the next nine days.
Table of Contents
- What Shipped
- The New CrawlForge CLI
- Extract With LLM: Local AI Extraction
- Scrape Template: Ten Sites, One Call
- list_ollama_models: Free Model Discovery
- Old Workflow vs v4.2.2 Workflow
- Credit Costs
- How to Upgrade
- What Is Next
What Shipped
v4.2.2 adds four things:
- @crawlforge/cli -- a standalone command-line tool exposing all 23 CrawlForge tools to your shell. No MCP client required.
- extract_with_llm -- LLM-powered structured extraction that defaults to local Ollama. No external API key needed.
- scrape_template -- pre-built scrapers for Amazon, LinkedIn, GitHub, YouTube, Reddit, Hacker News, Stack Overflow, npm, Product Hunt, and Twitter/X.
- list_ollama_models -- a free discovery tool that lists models on your local Ollama instance.
Tool count goes from 20 to 23. The CLI is brand new -- it is not a tool, it is a delivery channel.
+----------------+ +-------------------+ +----------------+
| Your Shell | <--> | @crawlforge/cli | <--> | CrawlForge |
| (cron, CI) | | (JSON in/out) | | API + Tools |
+----------------+ +-------------------+ +----------------+
^
No MCP handshake.
Just HTTPS + stdout.
The New CrawlForge CLI
The CLI is the shortest path from intent to scraped data. You install it once, set an environment variable, and every CrawlForge tool becomes a command:
npm install -g @crawlforge/cli
export CRAWLFORGE_API_KEY="cf_live_your_key_here"
crawlforge scrape https://example.com
crawlforge search "best MCP servers 2026"
crawlforge research "AI agent frameworks" --depth 3
Why does this matter? Because MCP is great for AI agents, but a lot of scraping work is not an AI agent task. It is a cron job. A CI step. A one-off pull from your terminal. For that, you want JSON on stdout that pipes into jq, not a JSON-RPC handshake.
MCP is optimized for AI agents picking tools dynamically. The CLI is optimized for humans typing commands and scripts piping JSON. Different shapes for different jobs: All three paths hit the same backend, share the same credit balance, and use the same API key.Why have a CLI when MCP already exists?
Workflow
Best fit
Claude/Cursor agent
MCP
Cron job
CLI
GitHub Actions step
CLI
One-off terminal
CLI
Server in a loop
Raw API
Read the complete CrawlForge CLI guide for the full command reference and real-world workflows.
Extract With LLM: Local AI Extraction
extract_with_llm is structured extraction powered by a language model. You hand it a URL and a schema, it gives you back JSON. The new part is that it defaults to local Ollama rather than calling OpenAI or Anthropic.
{
"url": "https://news.ycombinator.com/item?id=123456",
"schema": {
"type": "object",
"properties": {
"title": { "type": "string" },
"points": { "type": "number" },
"comments": { "type": "number" }
}
},
"provider": "ollama",
"model": "llama3.1:8b"
}
Three things follow from the local-first default:
- No third-party API costs. The LLM is free. You only pay 3 CrawlForge credits per extraction.
- No data leaving your machine. Scraped content stays on localhost.
- No new API key to manage. If Ollama is installed, you are done.
Local models are great for predictable schemas (titles, prices, counts, ratings). For long-form reasoning -- summarizing a 10,000-word article, classifying nuanced sentiment, extracting fields that require world knowledge -- a frontier model still wins. Switch providers with one parameter: You pay the provider's per-token cost plus 3 CrawlForge credits. Same schema, same output shape.When to still use OpenAI or Anthropic
crawlforge extract https://example.com \
--provider anthropic \
--model claude-sonnet-4-6
Detailed guide: extract data with local LLMs.
Scrape Template: Ten Sites, One Call
scrape_template is for the long tail of scraping requests that all look the same: "get me product data from Amazon", "get me a GitHub repo's metadata", "get me the top posts on Hacker News today". You should not need to write CSS selectors for these. We did it once, we maintain it, you call it.
crawlforge template amazon --url "https://www.amazon.com/dp/B0CHX1W1XY"
crawlforge template github --url "https://github.com/anthropics/anthropic-sdk-python"
crawlforge template hackernews --top 10
Ten templates ship in this release:
| Template | What it returns | Credits |
|---|---|---|
amazon |
Product title, price, rating, reviews, images | 1 |
linkedin |
Profile name, headline, experience, skills | 1 |
github |
Repo metadata, stars, languages, README | 1 |
youtube |
Video title, views, channel, transcript | 1 |
reddit |
Post title, score, comments, top replies | 1 |
hackernews |
Story title, points, URL, comments | 1 |
stackoverflow |
Question, answers, accepted, vote counts | 1 |
npm |
Package metadata, weekly downloads, versions | 1 |
producthunt |
Product name, tagline, upvotes, makers | 1 |
tweet |
Tweet text, author, engagement, replies | 1 |
Full walkthrough with code: scrape Amazon, LinkedIn, and GitHub with one tool.
list_ollama_models: Free Model Discovery
Most useful as a sanity-check before running extract_with_llm. Lists every model on your local Ollama instance with name, size, and modified date.
crawlforge extract --list-ollama-models
Costs zero credits. It does no scraping, no LLM call -- it just queries Ollama's local API on 127.0.0.1:11434 and returns the result. If you have ever wondered which model you actually have installed, this is the answer.
Old Workflow vs v4.2.2 Workflow
| Task | Pre-4.2.2 | v4.2.2 |
|---|---|---|
| Scrape from your terminal | curl + custom parser, or boot a Node REPL | crawlforge scrape <url> |
| Extract structured data with LLM |
extract_structured (CSS selectors) or roll your own with Puppeteer + OpenAI |
extract_with_llm (Ollama default) |
| Scrape Amazon, LinkedIn, GitHub |
scrape_structured with hand-maintained selectors |
scrape_template (we maintain selectors) |
| Run scraping in CI/cron | curl with API key in headers |
crawlforge <cmd> with env var |
Credit Costs
The three new tools follow our existing credit-cost model. No surprises:
| Tool | Credits | Why |
|---|---|---|
list_ollama_models |
0 | Free discovery helper |
scrape_template |
1 | Single page, pre-built schema |
extract_with_llm |
3 | LLM inference (provider-agnostic) |
The CLI itself is free. It uses your existing API key and bills against your normal credit balance.
How to Upgrade
Existing users do not need to do anything. The new tools are live on all plans -- Free, Hobby, Professional, and Business -- and show up automatically in your MCP client.
Add the Install the CLI
npm install -g @crawlforge/cli
export CRAWLFORGE_API_KEY="cf_live_..."
crawlforge --help
export line to your shell profile (~/.zshrc, ~/.bashrc) so it persists. For CI, set CRAWLFORGE_API_KEY as a repository secret.
The first run pulls about 5 GB. After that, every extraction is local, free, and offline-capable.Try Ollama-powered extraction
# 1. Install Ollama (one-time)
curl -fsSL https://ollama.com/install.sh | sh
# 2. Pull a model (llama3.1:8b is a good start)
ollama pull llama3.1:8b
# 3. Run extraction through CrawlForge
crawlforge extract https://example.com \
--provider ollama \
--model llama3.1:8b
What Is Next
We are working on three things for 4.3:
- More templates -- Etsy, eBay, TikTok, Instagram, Google Maps. Send us requests on Discord.
- Webhook delivery for batch_scrape -- get results pushed to your endpoint when long-running jobs complete.
-
CLI watch mode --
crawlforge track --watchfor live diffs on monitored pages.
Ready to try the new tools? Free tier still includes 1,000 credits and no credit card.
Or jump straight into the deep dives:
Top comments (0)