What is pi-agent-browser-screenshot?

pi-agent-browser-screenshot is an extension for a terminal based AI coding tool called pi. Its job is to let you see screenshots taken by another tool, agent-browser, directly inside pi's text interface, including when you are running pi inside tmux, a terminal multiplexer that normally blocks pi from showing images at all. When an AI agent using pi runs an agent-browser command that takes a screenshot, this extension notices the resulting image file and displays it right there in the conversation transcript. These screenshots only appear on your screen. They are saved along with the session but are never actually sent to the underlying AI model. How well this works depends on your terminal setup. In kitty or Ghostty terminals running inside tmux, the extension displays real inline images using a special graphics protocol built into those terminals. Outside of tmux, it falls back to whatever pi's own interface normally supports, which may still be an inline image or a plain text fallback depending on the terminal. In other terminal types running inside tmux, you only get a text fallback showing the file path instead of an actual image. To make inline images work inside tmux, you need tmux version 3.3 or newer, and you have to turn on a setting called allow-passthrough in your tmux configuration file, since tmux blocks the necessary escape sequences by default. If that setting is off, the extension simply falls back to showing text instead of failing. Technically, images inside tmux are transferred using the kitty graphics protocol, wrapped so it can pass through tmux, and drawn on screen using special Unicode placeholder characters that carry the image's identity and position as ordinary text. Because these placeholders behave like normal text, tmux can scroll and redraw them along with everything else. A few limits apply: only image files that the screenshot command itself just created are shown, images merely referenced elsewhere in the output are ignored, the tmux path only supports PNG images, and any file larger than 10 megabytes is skipped. The project is released under the MIT license.