openclaw-rpa — Locale router (read this first)#

GitHub: https://github.com/laziobird/openclaw-rpa — source, README, install, sample scripts under rpa/.

What this skill does#

openclaw-rpa is an LLM-based RPA Agent framework. You describe a task in plain language; the AI executes it step by step in a real browser, on your computer, or via API services — with screenshot proof at every step — then compiles everything into a standalone Playwright Python script. Replay runs that script directly — no model call, no token burn, no hallucination risk — faster and cheaper than having the AI click every time.

Why this matters

1. Saves compute and money — Having a large model drive the browser on every run can cost roughly single-digit to tens of US dollars per heavy session (tokens, tools, long context). After you record once, repeat work does not call the model—replay is much faster and near-zero LLM cost for those steps.
2. Verify once, run the same way every time — During recording you confirm the flow works; later, replay executes the saved steps deterministically. You avoid asking the AI to “do it again” on every run, which hurts consistency and raises hallucination risk.

What you can automate (record once, replay many times — follow each site’s terms and local law):

CategoryExamplesBrowserLogin, navigate, click, fill forms, extract text, sort / filter tablesHTTP APIGET / POST any REST endpoint, save JSON, embed API keys directly in the script ( #rpa-api)Excel ( .xlsx)Create / update workbooks, multiple sheets, headers, freeze panes, dynamic rows from JSON or another fileWord ( .docx)Generate reports with paragraphs and tables — no Microsoft Office requiredAuto-loginSave cookies once with #rpa-login, auto-inject on every recording and replay — skip OTP / CAPTCHA / QR-code flowsMixed flowsAny combination above in a single recorded task (e.g. API + Excel + Word, or browser + login + extract)

---

🚀 Real-world cases — already recorded, ready to run#

These scripts are registered in registry.json. Run any of them instantly — no re-recording, no tokens.

text

#rpa-list                          ← see all registered tasks
#rpa-run:amazonbestseller          ← run one directly
python3 rpa_manager.py run <name>  ← or via CLI

#CaseWhat it doesRun name🛒**Amazon Best Sellers ScraperScrape top 40 products (title, price, rating, reviews, URL) → Word tableamazonbestseller🏨Airbnb Competitor Price TrackerOpen browser → vision recognition → extract prices & ratings → Word reportairbnb民宿比价分析v11🏦AP Reconciliation (EN)Mock GET open payables → Excel match → Word table reportreconciliationV2🔑Auto-login: Ctrip HotelSave cookies once → skip OTP forever → extract hotel info → Word doc携程酒店V3📈Yahoo Finance NewsSearch NVDA → News tab → save top 5 headlines to DesktopYahooNew🎬Douban MovieSearch a film → detail page → title, rating, synopsis → Desktop file获取豆瓣电影数据🌐Alpha Vantage API**TIME_SERIES_DAILY for NVDA → nvda_time_series_daily.json (no browser)apiV3🛍️Sauce Demo ShoppingSign in → sort by price → add two most expensive → log outonlineShoppingV1

📖 Full case gallery & videos → README.md

---

When to use#

You want to…SendStart recording a new flow#automation robot, #RPA, #rpa, or mention Playwright automation****See saved tasks you can run#rpa-listRun a saved task (e.g. new chat)#rpa-run:{task name}Run in this chatrun:{task name} ( zh-CN: #run:{task-name})

Quick start (after install)#

text

#rpa-list
#rpa-run:your-task-name

Full protocol, state machine, two-line signup (task name + capability A–G/N), deps-check / deps-install, record-step JSON, progressive probing, and selector strength (composite CSS — container + tag / attributes / :has(); avoid bare h3) live in the locale file below.

Output#

Generated file is ordinary Python ( rpa/*.py) — runs standalone with python3, editable, no OpenClaw dependency at replay time.

Runtime (OS)#

macOS ( darwin), Linux, and Windows ( win32). The skill metadata lists all three for marketplaces (e.g. ClawHub). Stack: Python 3 + Playwright + httpx / openpyxl / python-docx — all support Windows. Details: README.md → Platforms (runtime).

Scope#

Browser — goto, click, fill, select_option, scroll, wait, snapshot, extract_text, dom_inspect.

HTTP API — api_call (httpx GET/POST, key embedding, save_response_to); independent of the browser page.

Local files — merge_files (concatenate Desktop files); extract_text writes to disk; patch rpa/*.py for folder / file ops after recording.

Excel / Word — excel_write (openpyxl, multi-sheet, dynamic rows from JSON or another file); word_write (python-docx, paragraphs + tables); no Microsoft apps required.

Computed logic — python_snippet injects arbitrary Python into the generated script; executed and validated at record time.

Out of scope — large ETL, databases, heavy OS automation.

Recommended sites#

Good fits — predictable structure, works well out of the box:

CategoryExamplesFinance / dataYahoo Finance, investing.comE-commerceSauce Demo ( saucedemo.com), AliExpress, eBayNews / mediaBBC News, Reuters, Hacker News, Reddit listing pagesReferenceWikipedia, GitHub public repo / issues pages

Not recommended — likely to break or require manual intervention:

SituationWhyWorkaroundLogin-gated flows (password / SMS OTP / slider / QR code)Credentials and 2FA must be handled manuallyUse #rpa-login to log in once manually → cookies saved automatically → #rpa-autologin injects them on every future recording and replay, skipping the login flow entirely

Tip: on a new site, start with goto + snapshot to confirm the page structure is readable before building a full flow.

Mandatory: load the correct instruction file#

1. Read config.json in this skill directory. If it does not exist, read config.example.json (same shape; default locale is en-US).

2. Read the "locale" field. Allowed values: zh-CN and en-US (repository default in config.example.json: en-US).

3. Immediately use the Read tool to load the full skill body:

   • zh-CN → SKILL.zh-CN.md
   • en-US → SKILL.en-US.md

4. Follow only that file for state machine, triggers, record-step JSON, onboarding text, and user-facing replies.

5. Reply to the user in the active locale’s language:

   • zh-CN → Simplified Chinese for agent messages (user may still type English).
   • en-US → English for agent messages (user may still type Chinese).

Changing language#

• Copy config.example.json → config.json if needed ( python3 scripts/bootstrap_config.py), then edit "locale", or
• Run: python3 scripts/set_locale.py en-US / python3 scripts/set_locale.py zh-CN (creates config.json from the example when missing).

After a locale change, the agent should re-read the matching SKILL.*.md in a new turn or session. See README.md in this directory for the full workflow.

ClawHub / discovery#

• SKILL.md (this file): short router + when to use + quick start for listings like ClawHub.
• SKILL.zh-CN.md / SKILL.en-US.md: full onboarding, recording, RUN/LIST, and anti-timeout rules.
• Scenario docs: Amazon Best Sellers · Airbnb Price Tracker · AP Reconciliation EN · AP Reconciliation CN · Auto-login.

Relative paths#

When the loaded file references playwright-templates.md, synthesis-prompt.md, or rpa_manager.py, resolve paths relative to this skill directory (parent of SKILL.md).