Crawling for AI agents

Point your agent at any site and get back the pages most relevant to your topic — or any single page as clean Markdown.

Open source
(query to search)
6.5×
Fewer LLM tokens
MD
From any page
MCP
One-line connect
0
Config to start
MCP built in

Six tools for your agent

Connect once — your agent can crawl a site by topic, compare the frontier strategies, read any page as clean Markdown, pull structured data, list a page's links, and call hidden APIs.

focused_crawl

Crawl a whole site and get its pages ranked by relevance to a query.

compare_strategies

Run Shark-Search, OPIC and BFS on one budget and see which finds the topic.

fetch_markdown

Any page as clean Markdown — a fraction of the tokens of raw HTML.

extract

Structured records out of a page with CSS selectors.

list_links

Every outbound link, absolute and deduplicated — the pages you could visit next.

fetch_json_api

Hit the JSON API hiding behind a site and skip the HTML.

Focused crawling

Three strategies, one fair loop

Every site is a graph: pages are nodes, links are edges. With a limited budget, a crawler decides one thing — which node to visit next. Same loop in all three; only the ordering changes.

BFS

Level by level from the seed. Uniform coverage, blind to your topic — the fair baseline.

score = −depth

Shark-Search

Hersovici et al., 1998

Best-first by topic: links inherit their parent's relevance with decay, so barren branches die off on their own.

score = γ·inherited + (1−γ)·local

OPIC

Abiteboul et al., 2003

Every page holds "cash" it splits among its links when visited — a live PageRank, no full graph needed.

cash(page) → split across links
Get started

Build with ByteCrawl today

Connect it to your agent, curl the API, or import the library — then explore every strategy live.