Webclaw Reviews

I’m not a developer; I used the hosted API for business research: prospect briefs on convenience-store chains and a dataset of 78 store, assistant and district manager job postings from 53 employers, all for under 200 credits. Map and batch scrape were the standouts, and one map call surfaced thousands of job URLs for a single credit. Watch two things: search scrapes every result unless you set scrape to false (my first search cost about 39 credits), and big-chain career sites (QuikTrip, Casey’s, Workday and Paycom portals) come back blocked or empty, while LinkedIn public postings, iCIMS and Greenhouse read cleanly. I also hit the same transient 502 on search, and brand missed the logo and real colors on a c-store site, so check its output.

I started on Tier 2 and upgraded twice — Tier 3, then Tier 4 — purely because I wanted the headroom, not because I hit a wall. That's my only criticism of the pricing structure: the tiers are generous enough that I ended up paying 4x what I needed, and I regret not starting higher anyway. This is a vendor complaint I made with my own wallet, not with the product.

What I use it for: I run three separate AI agent machines and wired the MCP server into all of them in a single pass. It's the extraction layer under everything.

Pros:

One credit per page, full stop. Not "one credit for a plain fetch, more for JS, more for blocked pages, 4-5x for structured extraction." Rendering, antibot, and AI extraction are all included in the single credit. That pricing is what made it easy to adopt — I could tell my agents to use it freely without modeling cost per call.

The MCP server is the whole product for me. One npx -y @webclaw/mcp and14 tools. Scrape, crawl, map, batch, extract, summarize, diff, brand, search, research. list_extractors returns 30+ site-specific verticals (GitHub, Reddit, YouTube, npm, PyPI, Shopify, WooCommerce) that return typed JSON instead of markdown — github_repo gave me stars, license, topics, and push activity in about 1.3KB with no API call and no LLM.

The llm output format is the real win. Same page: built-in extraction gave me 8,060 characters; webclaw's format: llm gave 4,309; adding the query filter ("what services does the company offer") dropped it to 2,610. About 68% fewer tokens for the same content, and my agents run in long sessions where that is the difference between fitting and truncating.
extract and summarize work with a local Ollama model. I point OLLAMA_MODEL at a cloud Ollama model, so those tools cost me nothing extra and never touch my agent's own model budget. That's a genuinely clever design — the LLM tools are a separate provider chain (Ollama → OpenAI → Anthropic) that you point wherever you want.
brand returned my exact hex colors, logo URL, favicon, and OG image in one call. Straightforwardly useful.
crawl respects robots.txt and reports per-page errors. Crawled 3 pages in 1.7s, 3 ok, 0 errors, 0 disallowed.
The developer is the founder. Valerio writes the code and runs the servers, so when something breaks you're emailing the person who broke it. No ticket queue, no handoff to a support tier.

Cons:

The prebuilt MCP binary ships without browser support. This is the one real limitation and it surprised me. Locally, render: always fails with "this webclaw-mcp build has no browser support; build it with --features browser." So local extraction is plain HTTP only — JS rendering and Cloudflare solving only happen via the hosted API (which costs a credit, so still fine in practice, but it's a design decision you should know about going in).
Antibot verticals don't always escalate as documented. The amazon_product extractor returned all-null fields with data_source: "local" instead of escalating to the cloud the way the docs imply. I had to fall back to the API directly for that one. Worth verifying the payload is actually populated rather than trusting the extractor name.
map has no cap parameter. Pointed it at a mid-size marketing site and got 20,060 URLs / 1.1M characters in one response. There's no max_pages on map at all. If you're going to use it, size the site first, or use crawl with explicit depth and max_pages instead, which is what I'd recommend.
Cloud search returned a transient 502 on my first call, then 200 on retry. Probably just a blip, but worth a retry if it looks broken.
GET /v1/account and GET /v1/me return empty bodies. /v1/usage is the endpoint that actually works for tier and quota. Minor, but I lost a couple of minutes to it.

Bottom line: The one-credit-per-page pricing with rendering, antibot, and LLM extraction all included is unusual and it's the whole reason I upgraded twice — the tool that most vendors meter at5-10x is the normal case here. The MCP server is well built and the 30+ vertical extractors are a big time saver over hand-rolling scrapers. My one real request is a max_pages cap on map and either shipping browser support in the prebuilt binary or making the escalation behavior consistent. Both are fixable in a day. Happy to answer questions about my setup.