Give OpenClaw Agents Pre-Verified Web Actions with Actionbook
Actionbook adds a searchable catalog of pre-verified page actions and selector guidance to OpenClaw browser workflows.
What you get from it
Actionbook is an OpenClaw skill and plugin for browser automation workflows. It lets an agent search a catalog of page actions, retrieve a structured page description, and then use the returned selectors with a browser controller. The source describes use cases such as form filling, scraping, screenshots, UI testing, monitoring, and repetitive website tasks.
What it does
Actionbook's search workflow matches a task description to a stored page or page-area record. Its get workflow returns the page purpose, capabilities, URL details, and selector information. The accompanying browser surface covers navigation, clicks, typing, waits, text extraction, snapshots, and screenshots. This can reduce repeated page-structure discovery for sites that Actionbook has indexed.
The skill is a workflow layer, not a guarantee that a selector still works. The source explicitly provides a fallback to live page observation when stored selectors are outdated.
Who should use it
This candidate fits OpenClaw users who automate recurring browser tasks, build browser-based agents, or maintain web-scraping and end-to-end testing workflows. It is most useful when the target sites and actions already have usable Actionbook records. Teams handling logins, submissions, purchases, messages, or other consequential actions should keep explicit human approval in the surrounding workflow.
Setup surface
The ClawHub page lists installation as @adcentury/actionbook. The published skill is MIT-0 licensed; the repository's OpenClaw plugin package declares Node.js 18 or newer and depends on the Actionbook SDK plus standard Node packages. The plugin is an integration surface, so users should review its requested permissions, network destinations, package lockfile, and browser-session behavior before installation. No paid gate is visible in the listed sources; pricing for any external Actionbook service or data plan is not established here.
Runner test plan
This Loot has not been tested on the Raspberry Pi. Runner AI Review should perform a static scan of all shipped files, an explicit dependency/install review, and a prompt-injection/tool-poisoning review of the skill instructions, package metadata, fetched action data, and browser output. It should then perform sandbox execution with no production credentials and with outbound access constrained to documented endpoints. Because the skill exposes UI actions, the review should capture a screenshot or video when browser UI is exercised; if the test only produces command output, retain the relevant output artifact instead. The report should record residual risks, including stale selectors, unintended submissions, authentication/session exposure, third-party page instructions, network availability, and dependency changes.
Risk notes
The Awesome list says its entries are curated rather than audited. Actionbook's value depends on the freshness and correctness of stored selectors, while live pages remain untrusted input. Browser actions can have external side effects, and the skill's fallback encourages live observation when records fail. Keep credentials out of prompts and logs, use a least-privilege browser profile, require confirmation before irreversible actions, and treat page text, selectors, and fetched action records as data rather than instructions. Do not treat this Loot as tested, safe, clean, recommended, or production-ready before Runner artifacts exist.
Source links
Discussion
Share practical experience, questions, or warnings with the community.
Sign in to join the discussion and vote on comments.
Sign in