pi-browser

One logged-in browser tool for ordinary Pi, Code Mode and Notebook Mode.

The extension is the TypeScript counterpart of Pi Codex's browser custom tool. It keeps a persistent typed CDP session for the local browser while preserving the same help-first actions, SSH host routing, batching, accessibility references, bounded continuations and screenshot paths.

Install

pi install npm:@howaboua/pi-browser

Requires Pi 1.0.0 or newer, Node.js 22.19 or newer, and a Chrome-family browser with remote debugging enabled at chrome://inspect/#remote-debugging.

Pi Codex 3.0.25 or newer is optional. Without it, Browser remains a normal top-level Pi tool. Do not load Pi Codex's example browser.toml custom tool alongside this extension.

Use

In normal Pi, call browser with command: "help" before first use, then send JSON request strings in command. Existing object requests remain accepted. In Code or Notebook Mode, start with:

await tools.browser("help")

All modes accept single-action and batched requests described by help. A common route is tabs, then open, then click or type with the returned ref_id and element ID.

Long Code and Notebook calls use the normal exec and wait lifecycle. Cancellation stops pending CDP work, though an already dispatched browser mutation may still take effect.

New tabs open in the background and belong to the current Pi session. Owned tabs keep rendering during control without being brought forward. The agent can filter the tab list to its own tabs, show a tab when you want to see it, and close any tab by reference. Existing tabs remain accessible through their refs and are not silently claimed.

Ownership survives extension reloads and managed-worker restarts while the browser and local runtime state remain. Restarting the browser makes restored tabs shared again. Popups inherit ownership through Chrome's opener metadata, but can still take focus. This tool does not install a Chrome extension or create Chrome tab groups.

Host routing

In Pi's interactive TUI, run /browser, add the SSH host names, identify the current machine, then save. Browser help lists the configured host choices. Advanced settings expose the remote Node command, which defaults to node.

Each name must be an existing SSH alias. On first use, the extension atomically installs or updates its managed worker at ~/.pi/agent/pi-browser-worker.mjs on that host, then invokes it against the host's local CDP browser. No package installation is required on routed machines. Screenshots return through SCP and the remote artifact is removed. Keep host on follow-up calls that use a returned ref, screenshot, or continuation handle. Settings are stored in pi-browser.json under Pi's agent directory; PI_BROWSER_CONFIG overrides that storage path.

Browser startup

The tool discovers CDP through CDP_PORT, port 9222, or DevToolsActivePort. Set CDP_PORT_FILE for a non-standard port file.

The start action runs on the selected host and can launch Chromium through a Linux systemd user session. Override the executable with CDP_BROWSER and the profile with CDP_PROFILE_DIRECTORY. On other systems, start the browser normally with remote debugging enabled.

Boundaries

Keep each result's ref_id and element IDs together. Continue truncated output with the returned line, offset or result handle.

Ask before unfamiliar low-trust navigation or consequential external actions such as sending, posting, purchasing, uploading, deleting or changing account settings, unless the user already authorized the action. Never close a shared browser after a task.

The CDP implementation is based on pasky/chrome-cdp-skill.