webtool is a fast, zero-dependency agent-first CLI that drives your Chrome browser.

webtool lets your agent connect directly to your browser using CDP. It does not require Playwright, a cloud browser, or a separate browser installation.

Just let your agent control your live browser session. webtool doesn't trigger bot detection because it is driving your real browser. It just works.

Your agent gets LLM-optimized semantic snapshots, token-efficient Markdown, and simple commands to click, type, and navigate.

What about security? It can be unsettling to give an AI agent full access to your browser. webtool has a powerful security policy engine that filters requests at the network level. Lock down the agent to specific pages or limit access with fine-grained request filtering. See docs/SECURITY.md for details.

npm i -g @usewebtool/webtool

webtool does not require Node.js, but npm is the easiest path to a cross-platform install.

npx skills add usewebtool/webtool

Install from source with Go:

go install github.com/usewebtool/webtool@latest

Or download a binary from the releases page and put it somewhere in your PATH.

To manually install the skill, clone the repo and copy skills/webtool to your agent's skills directory. For example, to install the skill into OpenClaw:

git clone https://github.com/usewebtool/webtool.git
cp -r webtool/skills/webtool ~/.openclaw/skills/

Make sure you're on the latest version of Chrome. You'll need to enable remote debugging so webtool can connect to your browser.

1. Open Chrome, navigate to chrome://inspect/#remote-debugging and enable remote debugging.
2. Start the webtool background process:

webtool start

3. Chrome will show a permission dialog. Click Accept.

That's it. webtool keeps the connection open, so you only approve once per session.

To stop webtool and close the connection:

webtool stop

Once you have installed the agent skill, just ask your agent to do things online.

"Open my Gmail and archive invitation emails."

The LLM is the brain. webtool gives your agent the "hands" it needs to interact with the web. Works in Codex, Claude Code, OpenClaw, or any other agent that supports skills.

You can use webtool from the command line or in shell scripts to programmatically control Chrome.

webtool start                       # connect to Chrome
webtool open https://example.com    # navigate to a URL
webtool snapshot                    # text snapshot of interactive elements
webtool click 43821                 # click an element by its ID from the snapshot
webtool type 43822 "hello world"    # type into an input field
webtool key Enter                   # press a key
webtool extract --main              # extract the main content as markdown
webtool tabs                        # list open tabs
webtool tab 2                       # switch to tab 2
webtool stop                        # close the connection 

See docs/usage.md for the full command reference.

You can create a simple YAML security policy file to filter your agent's Chrome traffic at the network level. Create a policy file:

# Read-only mode: block non-idempotent methods
deny:
  - method: "POST|PUT|DELETE|PATCH"

Or block all requests to specific sites. Wrap URLs in * wildcards to match all pages on the domain.

deny:
  - url: "*mail.google.com*"
  - url: "*bank.example.com*"
  - url: "*admin.example.com*"

Then start the daemon with it:

webtool start -p policy.yml

See docs/SECURITY.md for the full policy format.

Instead of feeding raw HTML or screenshots to your agent, webtool generates compact semantic snapshots. A page that would be 50k+ tokens as HTML becomes a few hundred tokens. The snapshots go beyond simple accessibility trees and create a structured map that is naturally easy for an LLM to understand.

Generate a snapshot of the current tab:

webtool open https://mail.google.com
webtool snapshot

[url] https://mail.google.com/mail/u/0/#inbox
[title] Inbox - Gmail

[1] navigation "Main"
  [2] link "Inbox" url="/mail/u/0/#inbox"
  [3] link "Starred" url="/mail/u/0/#starred"
  [4] link "Sent" url="/mail/u/0/#sent"
[10] list "Messages"
  [11] listitem "Alice Chen | Meeting tomorrow"
    [12] checkbox "Select"
    [13] link "Meeting tomorrow - Hey, are we still on for..."
  [14] listitem "GitHub | New issue assigned"
    [15] checkbox "Select"
    [16] link "New issue assigned - You've been assigned #421..."

See docs/FAQ.md.