web

name	web
description	Universal web tool — fetches and interacts with web pages through real Chromium (Playwright). Renders JavaScript, follows redirects, handles cookies, bypasses many anti-scrape stubs that block plain `web_fetch`. Each call launches a fresh browser, executes the requested flow, closes. ~2-3s cold start; no persistent state. Use INSTEAD of `web_fetch` when: - Target is a JS-heavy SPA (prices via XHR, lazy content) - `web_fetch` returned an anti-scrape placeholder (Zhihu's "click here" stub, JD's chrome-only HTML) - You need to interact (click, fill input) before extracting Common patterns: fetch rendered text: url=X wait for content then read: url=X wait_for=".price" extract specific element: url=X selector=".product-list" click and read result: url=X click=".search-button" fill form, then read: url=X click="input[name=q]" type_text="kinclaw" fill + Enter (React forms): url=X click="input" type_text="hi" press_enter=true wait_for=".reply" screenshot the page: url=X screenshot=true (returns image:// marker) run JS and get result: url=X js="document.title" First-time setup (one-time per machine, ~500MB Chromium download): pip install playwright playwright install chromium
command	["sh","-c","URL=\"$1\"\nWAIT_FOR=\"$2\"\nSELECTOR=\"$3\"\nCLICK=\"$4\"\nTYPE_TEXT=\"$5\"\nSCREENSHOT=\"$6\"\nJS_CODE=\"$7\"\nTIMEOUT=\"$8\"\nPRESS_ENTER=\"$9\"\n# Shifted positional params (>9) — POSIX shell needs ${} bracketing.\nSHOT_SELECTOR=\"${10}\"\nSHOT_FULL_PAGE=\"${11}\"\nPDF=\"${12}\"\nPDF_FORMAT=\"${13}\"\nSESSION_ID=\"${14}\"\nUPLOAD_SELECTOR=\"${15}\"\nUPLOAD_FILES=\"${16}\"\n\n[ -z \"$URL\" ] && { echo \"url required\" >&2; exit 1; }\n\n# Build args array — only include flags that have values.\nset -- \"$URL\"\n[ -n \"$WAIT_FOR\" ] && set -- \"$@\" --wait-for \"$WAIT_FOR\"\n[ -n \"$SELECTOR\" ] && set -- \"$@\" --selector \"$SELECTOR\"\n[ -n \"$CLICK\" ] && set -- \"$@\" --click \"$CLICK\"\n[ -n \"$TYPE_TEXT\" ] && set -- \"$@\" --type-text \"$TYPE_TEXT\"\n[ -n \"$JS_CODE\" ] && set -- \"$@\" --js \"$JS_CODE\"\n[ -n \"$TIMEOUT\" ] && set -- \"$@\" --timeout-ms \"$TIMEOUT\"\n[ \"$PRESS_ENTER\" = \"true\" ] && set -- \"$@\" --press-enter\n[ -n \"$SHOT_SELECTOR\" ] && set -- \"$@\" --screenshot-selector \"$SHOT_SELECTOR\"\n[ \"$SHOT_FULL_PAGE\" = \"true\" ] && set -- \"$@\" --screenshot-full-page\n[ -n \"$SESSION_ID\" ] && set -- \"$@\" --session-id \"$SESSION_ID\"\n[ -n \"$UPLOAD_SELECTOR\" ] && set -- \"$@\" --upload-selector \"$UPLOAD_SELECTOR\"\n[ -n \"$UPLOAD_FILES\" ] && set -- \"$@\" --upload-files \"$UPLOAD_FILES\"\n\n# Screenshot mode: generate output path under cache dir, pass it in.\nif [ \"$SCREENSHOT\" = \"true\" ]; then\n OUT=\"$HOME/Library/Caches/kinclaw/web/web-$(date +%Y%m%d-%H%M%S-%N).png\"\n mkdir -p \"$(dirname \"$OUT\")\"\n set -- \"$@\" --screenshot \"$OUT\"\nfi\n\n# PDF mode: same idea — generate output path, pass to --pdf.\nif [ \"$PDF\" = \"true\" ]; then\n OUT=\"$HOME/Library/Caches/kinclaw/web/web-$(date +%Y%m%d-%H%M%S-%N).pdf\"\n mkdir -p \"$(dirname \"$OUT\")\"\n set -- \"$@\" --pdf \"$OUT\"\n [ -n \"$PDF_FORMAT\" ] && set -- \"$@\" --pdf-format \"$PDF_FORMAT\"\nfi\n\n# Fail fast if playwright isn't installed — clear hint instead of\n# cryptic Python ModuleNotFoundError reaching the agent.\nif ! python3 -c \"import playwright\" 2>/dev/null; then\n echo \"playwright not installed — first-time setup:\" >&2\n echo \" pip install playwright\" >&2\n echo \" playwright install chromium\" >&2\n exit 1\nfi\n\n# cwd is already the skill directory (set by kinclaw), so a plain\n# relative path is reliable regardless of how SKILL_DIR is resolved.\npython3 ./web.py \"$@\"\n","_"]
args	["{{url}}","{{wait_for}}","{{selector}}","{{click}}","{{type_text}}","{{screenshot}}","{{js}}","{{timeout_ms}}","{{press_enter}}","{{screenshot_selector}}","{{screenshot_full_page}}","{{pdf}}","{{pdf_format}}","{{session_id}}","{{upload_selector}}","{{upload_files}}"]
schema	{"url":{"type":"string","description":"Target URL to load. Required.","required":true},"wait_for":{"type":"string","description":"Optional CSS selector to await visibility before extracting (e.g. \".price\", \"#main\"). Useful for SPAs where content arrives via XHR."},"selector":{"type":"string","description":"Optional CSS selector for the element whose inner text to return. Default body (full page text)."},"click":{"type":"string","description":"Optional CSS selector to click after load, before extracting. Combined with type_text, fills an input."},"type_text":{"type":"string","description":"Text to fill into the click target (only used with click=). Useful for search boxes."},"screenshot":{"type":"string","description":"Set to \"true\" to return a PNG of the viewport (with image:// marker for vision-capable brains). When true, no text is returned."},"js":{"type":"string","description":"JavaScript expression to evaluate in the page context. Result is JSON-stringified. When set, no other extraction happens."},"timeout_ms":{"type":"integer","description":"Max ms for goto and wait operations. Default 15000."},"press_enter":{"type":"string","description":"Set to \"true\" after fill to send a real Enter keypress on the\ninput — works for React/Vue/Svelte chat forms whose submit\nbutton is gated on internal state and won't enable from a fill\nalone. Real keyboard event (page.press), not synthetic. Use\nwith click=<input selector> + type_text=<message>.\n"},"screenshot_selector":{"type":"string","description":"When set with screenshot=true, capture only the bounding box of\nthis CSS selector instead of the full viewport. Falls back to\nviewport if the selector is offscreen / zero-sized. Useful for\ngrabbing just a chart, just a card, just a paragraph.\n"},"screenshot_full_page":{"type":"string","description":"When \"true\" (and screenshot=true, no screenshot_selector),\ncapture the FULL scrollable page rather than just the viewport.\nLong articles / infinite-scroll content in one shot.\n"},"pdf":{"type":"string","description":"Set to \"true\" to render the page as a PDF instead of HTML\nextraction. Mutually exclusive with screenshot=true. Headless\nChromium's print-pdf engine — high quality, includes JS-rendered\ncontent + computed CSS + background images. Default A4.\n"},"pdf_format":{"type":"string","description":"Page format for PDF mode (A4, Letter, Legal, A3, etc.). Default A4.\n"},"session_id":{"type":"string","description":"When set, persist storage state (cookies, localStorage) under\n~/.kinclaw/web-sessions/<session_id>.json. First call writes;\nsubsequent calls with same session_id load + reuse. Lets a\nsoul log into a site once and run multiple fetches without\nre-authenticating. Pair with --click=<login form> +\n--type-text=<creds> on the first call.\n"},"upload_selector":{"type":"string","description":"CSS selector for an <input type=file> element. Pair with\nupload_files= to populate the input via Playwright's\nset_input_files (fires change events the page expects). For\nhidden file inputs (drag-drop overlays), most still work via\nthe underlying input element.\n"},"upload_files":{"type":"string","description":"Comma-separated absolute file paths to upload via\nupload_selector. Examples: \"/Users/me/photo.jpg\" or\n\"/tmp/a.pdf,/tmp/b.pdf\" for multi-file inputs.\n"}}
timeout	60

web — universal Playwright-backed page tool

Single skill, single Python script (web.py), single Chromium per call. Drop-in replacement for web_fetch when JS rendering / dynamic content / minor interaction is needed.

Why one cold start per call is fine

Most web tasks the agent runs are one-shot: "fetch this page", "click this and read the result", "screenshot this URL". Cold start ~2s amortizes over a single useful task. Multi-step flows where state needs to persist (login → navigate → fill form → submit → read) are rare in agent usage; if you need them, write the whole flow as one web call by chaining click=... type_text=... wait_for=....

Architecture

agent → kinclaw shell skill → python3 web.py <args>
                              → Playwright launches Chromium
                              → executes flow
                              → closes Chromium
                              → prints result
                              → exits

No sidecar, no port management, no lifecycle skill. Just pip install playwright && playwright install chromium once per machine, then web works whenever the agent needs it.

What it doesn't do

• Persistent session (cookies, login state) across calls — each call is a fresh browser. For real-account tasks, drive the user's actual Safari via osascript activate Safari + ui skill.
• Anti-bot bypass on Cloudflare / DataDome / advanced fingerprinting — Playwright defaults are detectable. Closed-source web_scrape via WEB_ENDPOINT env is the answer when this matters.
• Sub-agent autonomy — this skill executes literal instructions, it doesn't plan multi-step research. For "find me cheapest X" the agent itself drives the loop (web_search → pick URLs → web fetch each → synthesize).