On this page
User guide
Browser
The agent drives a real Chromium (via Playwright) — not a text scraper. It can navigate, read rendered pages, click, fill forms, take screenshots, go back, and download files. Logins persist in the browser profile, so signing in once keeps working across runs.
Tools
| Tool | What it does |
|---|---|
browser_goto | Navigate to a URL (http/https only; credentials-in-URL rejected) |
browser_read | Read the rendered page as text |
browser_click | Click an element (by selector or visible text) |
browser_fill | Fill a form field |
browser_screenshot | Capture a screenshot to a file |
browser_back | Go back in history |
browser_download | Download a file to a chosen directory (no path collisions) |
The session starts lazily on first use and stays alive for the run. Pass headless=False when constructing a BrowserSession in code to watch the agent work in a visible window.
Setup
Playwright needs its Chromium build:
pip install playwright
playwright install chromium
Without it, browser tools raise a clear BrowserUnavailable error — they never pretend to work.
Safety notes
- Only
http/httpsURLs; anything else is rejected. - Downloads land in a directory you choose; filenames are de-duplicated, never overwritten silently.
- Like every tool, browser actions go through the permission gate: navigation/reading are low risk, form fills and downloads ask first.
- The agent does not bypass CAPTCHAs or access controls for you, and it won't exfiltrate credentials — login cookies stay in the local profile directory (
~/.agentkai/browser-profile/).
Limits
- JavaScript-heavy single-page apps sometimes need a wait/retry; the tools surface timeouts as errors rather than hanging.
- No built-in proxy support in v1; the browser uses your machine's network directly.