pi-agent-browser

Browser automation tool for Pi. Gives the model a browser tool that drives a real browser via agent-browser.

Install

pi install npm:@victor-software-house/pi-agent-browser

Or try without installing:

pi -e npm:@victor-software-house/pi-agent-browser

What it does

Registers a browser tool the model can call to:

• Navigate -- open <url>
• Inspect -- snapshot -i (interactive elements with @ref handles)
• Interact -- click @e1, fill @e2 "text", press Enter, scroll down
• Read -- get text, get title, get url, get text @e3
• Screenshot -- image returned inline so vision models can describe what they see
• Clean up -- close (also auto-closes on Pi session shutdown)

Any valid agent-browser subcommand works. The model passes the command string without the agent-browser prefix.

Features

| Feature | Details | |---|---| | Inline screenshots | Screenshots returned as base64 images for vision models | | Output truncation | Large snapshot output is truncated; full output saved to a temp file | | Session isolation | Each Pi session gets its own browser session (no cross-session collisions) | | Serialized execution | Browser calls are queued to prevent parallel tool-call race conditions | | Session cleanup | Browser auto-closed on Pi session shutdown | | TUI rendering | Compact display: element counts for snapshots, screenshot paths, errors |

Example

You: Open hacker news and tell me the top 3 stories

browser open https://news.ycombinator.com
browser wait --load networkidle
browser snapshot -i
browser close

The top 3 stories on Hacker News right now are:
1. ...

Operator commands

| Command | Description | |---|---| | /browser:doctor | Check binary, version, Chrome status | | /browser:examples | Print common workflow patterns |

Requirements

• agent-browser installed and on PATH:

  npm install -g agent-browser
  agent-browser install   # downloads Chrome

• A vision-capable model for screenshot descriptions (Claude Sonnet/Opus, GPT-4o, Gemini Pro)

License

MIT