Loading ToolBench
An MCP server that gives Claude web scraping powers
The server has 21 tools with descriptions present for most, but exhibits significant quality gaps. The three detailed tools (find_places, find_places_nearby, fetch_page) have strong, context-rich descriptions (150-250 chars each), explaining WHAT, WHEN, and HOW. However, 12 tools (extract_emails, extract_phones, extract_links, extract_social_links, extract_products, extract_by_selector, discover_sitemap_urls, crawl_by_links, detect_platform, fetch_shopify_collections, fetch_shopify_products, fetch_shopify_variants) have minimal descriptions under 50 characters, far below the 72-char minimum baseline for parameter annotations. Input schemas are visible for the three detailed tools with proper type definitions and descriptions for all parameters (find_places: 3 params with types and descriptions; find_places_nearby: 5 params; fetch_page: 7 params). However, 15 tools show NO visible input schema in the provided source, only function signatures. Error handling is present (try/except in find_places and find_places_nearby returning JSON with 'error' and empty results), but error messages lack recovery guidance. Parameter naming follows verb_noun convention (find_places, fetch_page), which is good. However, sensitive tools like make_request lack any permission gates or security annotations. Output structure is returned as JSON strings with query/error/count/results fields for the detailed tools, but 15 other tools have no documented output schemas. The rubric baseline is 4 params/tool average (p10=1, p90=8); the detailed tools (3-7 params) fall within range, but schema quality for the majority is invisible.
Crawl a site by following links
Delete a saved result table
Detect the platform or CMS of a website
Discover page URLs from a site's sitemap
Extract arbitrary fields from a web page using CSS selectors
Extract email addresses from a web page
Extract hyperlinks from a web page
15 of 21 tools lack visible input schemas in source code. This blocks LLM understanding of valid parameters.
12 tools have descriptions under 50 characters: extract_emails, extract_phones, extract_links, extract_social_links, extract_products, discover_sitemap_urls, crawl_by_links, detect_platform, fetch_pages, screenshot_page, list_tables, delete_table. Descriptions provide no context for tool selection or prerequisites.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 37 | <=2025-11-25 | v2 |
Extract phone numbers from a web page
Extract structured product data from a web page
Extract social media links from a web page
Fetch one web page and return its content as clean markdown with metadata. Set js_render=true for pages that need JavaScript to render (SPAs, infinite-scroll listings, most modern storefronts). If a rendered page still comes back sparse, give it longer with wait_ms, or wait for a specific element with wait_for (a CSS selector) — that is more reliable than a fixed delay. Article-shaped pages have their navigation, cookie banners and footers stripped automatically; set main_content=false to keep the whole page. PDFs are extracted to text. `actions` drives the page before reading it (implies js_render). Each is {action, selector, ...}: click, type (text), press (key), wait (ms), wait_for, scroll (times), scroll_to_bottom (max_rounds) for infinite scroll, and click_until_gone (max_clicks) for a 'load more' button. Use it for cookie gates, paginated listings and search forms.
Fetch many web pages in parallel
Fetch collections from a Shopify store
Fetch products from a Shopify store
Fetch product variants from a Shopify store
Search for places and local businesses by name or description — 'the British Museum', 'cafes in Shoreditch'. Returns name, address, coordinates, phone, website, opening hours and category. Data comes from OpenStreetMap, so there are no star ratings or review counts; for those you would need a paid Google Places key. Pass save_as='<table_name>' to store the results.
Find every business of a category within a radius of a place — 'dentists near Bath', 'cafes within 2km of Shoreditch'. This is the tool for local lead generation: it returns name, address, phone, website and opening hours for each. `category` accepts friendly names (cafe, dentist, hotel, solicitor, gym, hairdresser, ...) or a raw OpenStreetMap tag like 'amenity=dentist'. Data is OpenStreetMap, so there are no star ratings. Pass save_as='<table_name>' to store the results.
List saved result tables
Make custom HTTP requests
Query a saved result table with SQL
Take a screenshot of a web page
delete_table is a destructive tool with no confirmation mechanism, permission gate, or error classification. Agents can delete result tables without safeguards. Per pattern:confirmation-request, irreversible operations should support dry-run or confirmation.
make_request accepts arbitrary HTTP calls with no visible schema or input validation. LLM could pass malicious payloads or secrets as parameters. Per pattern:secret-injection, credentials must never be tool parameters; per pattern:tool-gateway, all input must be sanitized.
Output schemas are documented only for find_places, find_places_nearby, and fetch_page. 18 other tools have no visible output schema declaration. LLMs cannot plan downstream tool calls or extract required fields.
fetch_page has 7 parameters (js_render, wait_ms, wait_for, main_content, actions, as_html, plus implicit url) but 'actions' parameter is described as 'List of actions...' without specifying the action object schema (required fields, enum of valid actions: click, type, press, wait, wait_for, scroll, scroll_to_bottom, click_until_gone). LLM cannot construct valid action objects. Per pattern:constrained-input, enums must be explicit.
No tool annotations visible (readOnlyHint, destructiveHint, idempotentHint). Per current MCP spec (2026-07-28), tool annotations are a CURRENT pattern for marking tool safety properties. Only delete_table should be marked destructive; all others are read-only. Missing annotations reduce agent planning safety.
Error handling in find_places and find_places_nearby returns JSON with 'error' field but provides no recovery guidance. Per pattern:recovery-guide, errors must tell LLM what to do next. Generic PlacesError exceptions are caught but not classified as retryable or user-fixable.