Why your AI assistant needs a real browser
Modern AI coding assistants — Claude Code, Cursor, OpenClaw, Hermes, Codex, Gemini, Windsurf — are great at writing code, but they're blind to the web. When you ask Cursor to "scrape the YC company list" or Claude Code to "check what's new on Hacker News and summarize it", they either refuse, hallucinate the answer, or try to fetch raw HTML with a basic curl-style tool that fails on every modern JavaScript-rendered site.
Browse Anything fixes that gap by being a tool the assistant can call. The ClawHub skills protocol exposes Browse Anything as a typed capability your agent can invoke whenever it needs to interact with the web: navigate, click, type, scrape, screenshot, log in, fill forms — anything you'd do in a real browser, your AI assistant can now do too.
Crucially, this works in both directions. You can ask the assistant to use Browse Anything ("scrape that page for me"), or you can use Browse Anything's prompt language directly inside the assistant ("run this Browse Anything task and use the output to update my README"). The two layers compose naturally.
What ClawHub is and why it matters
ClawHub is the package manager for AI assistant skills — think 'npm for agent capabilities'. It standardizes how skills are discovered, installed, versioned, and invoked across a growing list of AI runtimes. The same skill installs identically into Claude Code, OpenClaw, Hermes, Cursor, and others, because each runtime implements the ClawHub skills protocol.
For you as a user, this means: install Browse Anything once on ClawHub, and it works everywhere. No per-IDE plugin to maintain. No three-way fork of the integration. Updates roll out via ClawHub the same way npm packages get new versions.
For Browse Anything specifically, the skill ships a curated set of high-leverage primitives — `browse`, `scrape`, `submit_form`, `screenshot`, `schedule_task`, `subscribe_to_changes` — that your assistant can compose freely. The skill is open and the manifest is public on ClawHub at https://clawhub.ai/mehdi149/browseanything.
Install in one command
Skill install through ClawHub is identical across runtimes — that's the whole point of standardization. Below are the exact commands for the most common assistants today.
- OpenClaw CLIRun `openclaw skills install browseanything`. Then in any project, the skill is available to whichever agent you start with `openclaw`.
- Claude CodeAdd Browse Anything to your `claude-skills.json` via `claude skills add browseanything` (or whatever Claude Code's current installer command is — check ClawHub docs). The skill autoloads on next session.
- CursorOpen Cursor settings → MCP / Skills → add the ClawHub Browse Anything entry with your API key. The agent can now call Browse Anything from any chat or composer session.
- HermesHermes consumes ClawHub skills natively. Install via the Hermes skill installer pointed at the ClawHub registry. Documented on the ClawHub page.
- Codex / Gemini / Windsurf / othersAny agent that implements the ClawHub skills protocol — or supports MCP — can install Browse Anything. Browse Anything also publishes a native MCP server for runtimes that prefer MCP over ClawHub directly.
What the skill lets your assistant do
Once installed, the skill exposes typed primitives that the host agent picks up automatically. The agent decides when to call them based on user intent; you don't have to tell it explicitly to use Browse Anything.
- browse(url, instructions)Navigate to a URL and follow plain-English instructions. Returns the final page state, extracted data, and a screenshot.
- scrape(prompt)One-shot scraping. The agent describes the data it wants and the skill returns structured JSON.
- submit_form(url, fields)Locate a form and submit it with the provided fields. Handles multi-step forms, dropdowns, file uploads, and consent banners.
- screenshot(url)Headful Chromium screenshot. Useful when the assistant needs to 'see' the page to reason about it (great for design review or QA work).
- schedule_task(prompt, schedule)Create a recurring task. The assistant can schedule reports, monitors, and pipelines without your hand-holding.
- subscribe_to_changes(url, fields)Set up a change-monitoring task. The assistant can wire up alerting workflows in one call.
Workflows that become natural once your IDE has a browser
These are the workflows users report after installing Browse Anything in their AI assistant. None of them require leaving the IDE.
- Read the docs the agent is implementing against"Read the Stripe webhook docs at /docs/webhooks and update my handler to support the new event types." The agent fetches the real page, reads it, edits your code accordingly.
- QA the app you just shipped"Open the staging URL, sign in as the test user, run the signup flow, screenshot any error." The agent does the full QA pass and reports back inline.
- Pull in production data the easy way"Log into Stripe and pull the count of subscriptions created last week, then update the dashboard query." Real data flows into your code review without context switching.
- Set up monitoring as a one-liner"Schedule a daily 9am check of competitor.com/pricing and alert me on Slack if anything changes." Skip the UI, the agent creates the scheduled task for you.
- Verify your changes affect the real web"Check whether my latest deploy is live at example.com — visit the page and confirm the new feature is rendered." Closes the feedback loop without leaving your editor.
- Generate datasets on demand"I need a CSV of the top 200 ProductHunt launches from 2025 sorted by upvotes — go get it." The agent calls Browse Anything, returns the file, and you keep coding.
How API keys, billing, and credits work
The skill is free to install. Each Browse Anything action invoked through the skill is billed to the Browse Anything API key you configure — including the free tier. If you only call the skill occasionally for one-off scrapes, the free tier covers it. If you run scheduled jobs or high-volume scraping from inside your IDE, the Pro ($9.99/mo) or Ultra ($45/mo) plan applies.
The skill respects your plan's quotas. If you exceed credits, calls return a clear error the assistant can communicate to you ("Browse Anything credits exhausted, upgrade at platform.browseanything.io") rather than failing silently or hanging. There's no double-billing — the ClawHub layer is free.