Web intelligence for AI agents.
Web scraping built for AI agents. Give your agents real-time web access: fetch any URL and get clean, token-budgeted, cached Markdown. Auto-routes JS-heavy pages and PDFs. One MCP tool, every URL type, fraction of a cent per call — a drop-in Firecrawl alternative for the agent stack.
500 free fetches. No credit card.
New · Official LangChain directory listing AgentFetch is available as an external integration in the official LangChain tools directory. Install the Python package and add token-budgeted web retrieval to your agent. View the LangChain integration →Try it
After signup we email your key. Then:
curl -X POST https://api.agentfetch.dev/fetch \
-H "Content-Type: application/json" \
-H "X-AgentFetch-Key: af_xxx" \
-d '{"url": "https://example.com",
"max_tokens": 2000}'Or use it as an MCP tool
Add to Claude Desktop / Claude Code:
{
"mcpServers": {
"agentfetch": {
"command": "python",
"args": ["-m", "agentfetch.mcp.server"],
"env": {"AGENTFETCH_KEY": "af_xxx"}
}
}
}Why agents prefer AgentFetch
Plain fetch() wastes 70–85% of your context window on nav,
footers, and scripts. AgentFetch is data-retrieval infrastructure for LLM agents that returns
LLM-ready Markdown instead — the fastest way to fetch web data for AI agents without writing a
scraper per runtime.
Token estimation
Call estimate_tokens before fetching. Avoid blowing your context window on a page that turns out to be 50K tokens.
Smart caching
6-hour cache by default. Repeat fetches are ~$0.0001 instead of $0.005.
Auto-routing
Trafilatura for simple HTML, Jina for the rest, FireCrawl for JS, pypdf for PDFs. You don't pick.
Per-call billing
No subscription required. From $0.001 per fetch. Top up in $10 increments.
Batch & concurrent fetch
Fetch multiple URLs for your AI agent in one call — fetch_multiple pulls up to 20 pages in parallel, so a research agent gathers a whole reading list in a single round-trip.
Search & fetch in one call
search_and_fetch runs a live web search and returns the clean Markdown of the top results together — real-time web data for LLMs without wiring up a separate search API.
Pricing
Free
$0
500 fetches/month. No credit card.
Credit packs
$10 / $80 / $600
10K / 100K / 1M credits. Never expire.
Subscriptions
From $29/mo
20K/100K/500K standard fetches included.
Frequently asked questions
How do I give my AI agents web access?
Add AgentFetch as an MCP server to Claude Desktop, Cursor, Cline, or your LangChain stack, or call the REST API directly. Your agent gets a fetch_url tool that returns clean, LLM-ready Markdown for any URL — no scraper code to maintain.
What's the best web scraping tool for AI agents?
For agent workloads, an MCP-native fetch layer beats raw HTTP or a general-purpose scraper: AgentFetch handles retries, rate limits, JS-heavy pages, PDFs, and HTML-to-Markdown conversion so the model spends tokens on content, not boilerplate.
Is AgentFetch a Firecrawl alternative?
Yes. AgentFetch covers the same fetch-and-extract job with an MCP-first install and per-call pricing from $0.001, plus 500 free fetches — a lightweight alternative to Firecrawl, Jina Reader, and ScrapingBee for the AI agent use case.
How does AgentFetch cut LLM token costs?
It converts raw HTML to clean Markdown (70–85% smaller), caches repeat fetches for six hours, and lets you estimate tokens before you fetch — so real-time web data reaches your LLM without blowing the context window.
How do I stop a big web page from blowing my agent's context window?
Call estimate_tokens on a URL first. AgentFetch returns the token count without pulling the full body, so your agent can skip or chunk a 50K-token page before it overflows the context window — and every fetch can be capped with max_tokens.
Can my AI agent search the web and fetch the results in one call?
Yes. search_and_fetch runs a live web search and returns the clean Markdown of the top results in a single round-trip, and fetch_multiple pulls up to 20 URLs in parallel — so you get real-time web data for your LLM without wiring up a separate search API.