Web scraping MCP
Web scraping MCP servers let your coding agent read web pages. We sent one Wikipedia article through all 6 and counted tokens: 94,415 to 112,517. FATSEARCH returned 1,384 in three reads and found the same answer.
We make FATSEARCH. The numbers below are our own runs, dated. Where another reader is the better pick, we say so.
| One page, | 38 hard pages, | |||||
|---|---|---|---|---|---|---|
| Reader | Tokens in context | Answer found | Model bill | Answer in the text | Agent answered right | Tokens per question |
| FATSEARCH | 1,384 | yes | $0.0028 | 35/38 | 26/38 | 2.6k |
| Jina Reader | 94,415 | yes | $0.19 | 8/38 | — | — |
| Tavily | 97,504 | yes | $0.20 | 19/38 | — | — |
| Oxylabs | 102,609 | yes | $0.21 | — | — | — |
| Bright Data | 103,625 | yes | $0.21 | — | — | — |
| Firecrawl | 112,517 | yes | $0.23 | 24/38 | 24/38 | 2.8k |
One page: Wikipedia “Rust (programming language)”, default settings, one run each. Tokens = Anthropic’s own counter; fats = outline + one filter + one section. Model bill = tokens × $2 per million input tokens (Claude Sonnet 5.5, ); reader fees not included.
Hard pages: Social, news, shops, travel, classifieds, search: pages a built-in fetch can’t read. FATSEARCH 0.1.0, 2026-10-16, one run each; Firecrawl’s copy was cut to fit the agent run. “—” = not in that run.
Each reader’s section below says when it’s the better pick. Method and raw data: benchmarks.
These are hosted readers. Open-source tools you run yourself went through the same page on : open source web search MCP, and the two crawlers head to head in Crawl4AI vs Firecrawl.
Comparisons
- Crawl4AI vs Firecrawl Crawl4AI: 74,896 tokens. Firecrawl hosted: 112,517, self-hosted: 107,332. FATSEARCH: 1,384. Same Wikipedia page.
- Firecrawl alternatives Firecrawl returned 112,517 tokens for one Wikipedia page. FATSEARCH got the same answer in 1,384. 6 readers, real counts.
- Jina Reader vs Firecrawl Jina Reader: 94,415 tokens. Firecrawl: 112,517. FATSEARCH: 1,384. Same Wikipedia page, real counts.
- Firecrawl vs Tavily Firecrawl: 112,517 tokens. Tavily: 97,504. FATSEARCH: 1,384. Same Wikipedia page, same answer, head-to-head.
- Jina Reader alternative Jina Reader returned 94,415 tokens for one Wikipedia page. FATSEARCH: 1,384, same answer. 6 readers with real numbers.
- Open source web search MCP One Wikipedia page through 6 open-source MCP tools, default settings: 2,373 to 117,088 tokens. FATSEARCH: 1,384 in three reads.
- Tavily alternatives Tavily returned 97,504 tokens for one Wikipedia page. FATSEARCH got the same answer in 1,384. 6 readers with real counts.
Guides
- Claude Code web fetch 403 Claude Code's WebFetch got a 403 from npm, Medium and G2. FATSEARCH read all three: npm in 345 ms for $0. Real output, dated.
- Web fetch vs web search Claude Code's two web tools on one question: WebSearch returned links (626 tokens), WebFetch a 302-token summary of the page.
- How to scrape Twitter How to scrape Twitter, priced: X's own API at $0.005 a post, Apify actors from $0.00025 a tweet, TwitterAPI.io at $0.15 per 1K tweets.
- RapidAPI scraper What a RapidAPI scraper listing really is: Nokia owns it, a 25% fee, no refunds, and dead APIs still sell. How to vet one before you pay.
- Web scraping MCP Claude Code Firecrawl MCP used 38,129 tokens per task in Claude Code, FATSEARCH 5,026. The same 51 tasks, Sonnet, headless, measured.
- Web search MCP for Claude Code Add a web search MCP server to Claude Code and check it works. Every step with its real output: FATSEARCH 0.1.3, Claude Code 2.1.280.
Reader by reader
FATSEARCH
Local CLI + MCP server, your own provider key. Measured: outline, then one filter, then one section (plain HTTP, $0).
1,384 tokens across three reads. First lines:
= j11.u46 Rust (programming language) - Wikipedia · en.wikipedia.org · 57 sections · 6 photos · 2 videos · 40.9ktok · pl… u46#0 info | author Contributors to Wikimedia projects · published 2010-10-30T22:30:54Z · modified 2026-10-15T20:56:59Z m1 photo 250x167 png | A bright orange crab icon → https://thumb.wikimedia.org/wikipedia/commons/thumb/2/20/Rustacean-or… m2 photo 250x314 png | (no alt) → https://thumb.wikimedia.org/wikipedia/commons/thumb/a/af/Rust_standard_libraries.svg/5… m3 photo 250x193 png | (no alt) → https://thumb.wikimedia.org/wikipedia/commons/thumb/6/6c/Cargo_clippy_hello_world_exam… … +10 media → fats u46 media u46#1 (top) (9) | Jump to content u46#9 Syntax and features (9.2k, 27 sub) | [edit [l175] ] Main article: Rust syntax [l176] Rust's syntax [l177] is simil…
Price: $0.001–0.0065 per hard page, $0 plain pages (github.com, checked )
Jina Reader
Hosted URL-to-Markdown reader, MCP server. Measured: r.jina.ai, no key.
94,415 tokens, 68× what FATSEARCH used. First lines:
Title: Rust (programming language) URL Source: https://en.wikipedia.org/wiki/Rust_(programming_language) Published Time: 2010-10-30T22:30:54Z Markdown Content: | Rust | | --- | | [ | * [Multi-paradigm](https://en…
Price: Free without a key at 20 requests a minute. With a key it is billed by output tokens; a new key comes with 10M free tokens. (jina.ai, checked )
When it’s better: Zero install, no account. Put r.jina.ai/ in front of a URL and you get Markdown back. Good for a quick one-off read in a script.
Tavily
Hosted search + extract API, MCP server. Measured: /extract, basic depth.
97,504 tokens, 70× what FATSEARCH used. First lines:
[Jump to content](#bodyContent) [   ## Contents * [1 History](#History) + [1.1 2006–2009: Early years](#2006–2009:_Early_years) + [1.2 2009–2012: Mozilla sponsorship](#2009–2012:_Mozilla_sponsorship) + [1.3 2012–2015: Evolution](#2012–2015:_Evolution)
Price: Free: 1,000 API credits a month. Pay as you go: $0.008 a credit. A basic extract costs 1 credit per 5 successful URLs (advanced: 2). (tavily.com, docs.tavily.com, checked )
When it’s better: Search is the main job. Tavily is a search API for agents and can return page text with the results in one call. FATSEARCH searches too, but hands back result ids your agent opens one by one.
Oxylabs
Hosted scraper API + proxies, MCP server. Measured: Web Scraper API /v1/scrape, markdown, no JS.
102,609 tokens, 74× what FATSEARCH used. First lines:
[Jump to content](#bodyContent) Main menu Main menu move to sidebar hide Navigation * [Main page](/wiki/Main_Page "Visit the main page [z]") * [Contents](/wiki/Wikipedia%3AContents "Guides to browsing Wikipedia")
Price: Web Scraper API, other sites: $1.00 per 1,000 results on the Starter plan ($30 a month). $5 of free welcome credits. (oxylabs.io, checked )
When it’s better: You need whole pages at volume, including Google and Amazon, with proxies and an unblocker under one subscription.
Bright Data
Hosted unblocker API + proxies, MCP server. Measured: Web Unlocker, markdown.
103,625 tokens, 75× what FATSEARCH used. First lines:
Rust (programming language) - Wikipedia [Jump to content](#bodyContent) Main menu Main menu move to sidebar hide Navigation * [Main page](/wiki/Main_Page "Visit the main page [z]") * [Contents](/wiki/Wikipedia:Contents "Guides to browsing Wikipedia")
Price: Web Unlocker: 5,000 free requests a month, then $1.5 per 1,000 requests, pay as you go. (brightdata.com, checked )
When it’s better: You need pages that fight back, at volume. Web Unlocker handles CAPTCHAs, browser rendering and retries. You pay only for success.
Firecrawl
Hosted scrape, crawl + search API (AGPL core), MCP server. Measured: v2 /scrape, markdown, default settings.
112,517 tokens, 81× what FATSEARCH used. First lines:
[Jump to content](https://en.wikipedia.org/wiki/Rust_(programming_language)#bodyContent) [ | - [Multi-paradigm](https://en…
Price: Free: 1,000 credits a month (one scraped page = one credit). Hobby: $16 a month billed yearly ($19 monthly) for 5,000 credits. Standard: $83 a month billed yearly for 100,000. (firecrawl.dev, checked )
When it’s better: You want a whole site, not one page. Firecrawl crawls a site in one request, maps its URLs, batch-scrapes many pages and returns structured JSON or screenshots. It can also search, click through a page, and turn PDFs and DOCX into Markdown. Its core is open source under the AGPL and can be self-hosted, though some features are Cloud-only. FATSEARCH has no crawler and no DOCX yet.
Questions
Which scraping service is the best?
Depends on the job. On 38 hard pages, the answer was in the text for FATSEARCH 35/38, Firecrawl 24/38, Tavily 19/38, Jina without a key 8/38. For site crawls, search plus text, or pages at volume, see each reader's section above.
Is Firecrawl MCP free?
The MCP server is open source (MIT). Firecrawl also hosts an MCP endpoint you can use without a key for search, scrape and parse, capped per IP address per day. Past that you need a key. Free: 1,000 credits a month (one scraped page = one credit). Hobby: $16 a month billed yearly ($19 monthly) for 5,000 credits. Standard: $83 a month billed yearly for 100,000. You can also self-host Firecrawl's AGPL core, but some features are Cloud-only.
Is there a free alternative to Tavily?
FATSEARCH is open source and reads plain pages for free. Search, and pages that need a provider, run on your own Oxylabs or Bright Data key: $0.001–0.0065 per hard page, $0 plain pages. Jina Reader: Free without a key at 20 requests a minute. With a key it is billed by output tokens; a new key comes with 10M free tokens. Tavily itself: Free: 1,000 API credits a month. Pay as you go: $0.008 a credit. A basic extract costs 1 credit per 5 successful URLs (advanced: 2).
Try FATSEARCH on your own pages
Free and open source. Bring any one key, or none.
curl -fsSL https://fatsearch.com/install.sh | shnpm i -g fatsearchbrew install fatsearch/tap/fatsearchcargo install --locked fatsearchor let your agent do it
Read and follow fatsearch.com/agent-setup/SKILL.md