English | 中文

A local-first browser rendering CLI for live preview, browser control, and DevTools capture. Write HTML/CSS/JS in a local directory and have it rendered in a real Chromium browser with full DevTools control. 🎨

AI agents that generate HTML/CSS/JS need to see what they built. They need to render the page in a real browser, take a screenshot, read the DOM, check for console errors, and iterate — all without a human clicking around.

Existing tools don't fit this workflow:

The typical AI agent workflow with browsercli:

Agent writes HTML/CSS/JS to disk
        ↓
browsercli auto-reloads the browser (250ms)
        ↓
Agent runs: browsercli screenshot --out page.png
        ↓
Agent inspects the screenshot / queries DOM / checks console
        ↓
Agent iterates on the code

No test framework boilerplate. No browser lifecycle management. Just a persistent browser that reflects your files and responds to commands.

• Daemon architecture — browsercli start launches a background process; CLI commands talk to it via Unix socket RPC (macOS/Linux) or TCP localhost (Windows)
• Static file server — Serves a local directory over HTTP with automatic index.html resolution
• Auto-reload — File watcher with 250ms debounce triggers browser reload on save
• App mode — Opens a chromeless (--app) browser window by default
• Stealth mode — Best-effort automation detection reduction (webdriver flag removal)
• Full DOM control — Query, query-all, attr, click, type, wait via CSS selectors
• JavaScript evaluation — Execute arbitrary JS in the controlled tab
• Screenshot capture — Full page or element-specific (PNG)
• Console capture — View browser console output (log, warn, error, info) with level filtering and --clear to drain the buffer
• Network logging — Inspect HTTP requests/responses with method, status, resource type, MIME type, size, and duration; supports --clear
• Performance metrics — Navigation Timing Level 2 (with legacy fallback) for DOMContentLoaded and Load timing
• JSON output — Pass --json for machine-readable output on every command
• Plugin system — Extend browsercli with custom page templates, RPC endpoints, and lifecycle hooks via script-based plugins with JSON manifests
• Cross-platform — macOS (app bundles), Linux, and Windows Chrome/Chromium/Edge auto-detection

Download the latest binary for your platform from GitHub Releases:

Extract the archive and place the binary in your $PATH.

brew tap justinhuangcode/tap
brew install browsercli

cargo install browsercli

# Node.js
npm install @justinhuangcode/browsercli

# Python
pip install browsercli

cargo install --path .

Requirements: Rust 1.75+ and a Chromium-based browser (Chrome, Chromium, Brave, or Edge). On Windows, Microsoft Edge works out of the box.

# Start with a project directory
browsercli start --dir ./my-site

# Or start with a temp directory
browsercli start

# Check status
browsercli status

# Navigate to a path
browsercli goto /

# Query the DOM
browsercli dom query "h1" --mode text
browsercli dom all "a" --mode outer_html

# Evaluate JavaScript
browsercli eval "document.title"

# Take a screenshot
browsercli screenshot --out page.png

# View console output (--clear drains the buffer)
browsercli console --level error
browsercli console --clear

# View network requests (--clear drains the buffer)
browsercli network --limit 20
browsercli network --clear

# Performance timing
browsercli perf

# Stop
browsercli stop

browsercli dom query "selector" [--mode outer_html|text]
browsercli dom all "selector" [--mode outer_html|text] [--limit N]
browsercli dom attr "selector" "attribute-name"
browsercli dom click "selector"
browsercli dom type "selector" "text" [--clear]
browsercli dom wait "selector" [--state visible|hidden|present|gone] [--timeout 10s]

Shorthand:

browsercli dom "#app" --mode text

browsercli has a built-in plugin system with three extension points: page templates, custom RPC endpoints, and lifecycle hooks. Plugins are plain directories with a plugin.json manifest and executable scripts -- no compilation, WASM, or dynamic libraries required.

~/.browsercli/plugins/my-plugin/
├── plugin.json              # Manifest (required)
├── templates/
│   └── dashboard/           # HTML/CSS/JS scaffold
│       ├── index.html
│       ├── style.css
│       └── app.js
├── handlers/
│   └── refresh.sh           # Custom RPC endpoint script
└── hooks/
    └── on_start.sh          # Lifecycle hook script

browsercli ships with 4 built-in templates that work out of the box — no plugins needed:

browsercli start --template tailwind
browsercli start --template dashboard
browsercli start --template chart
browsercli start --template form

All templates are single-file HTML with CDN imports — zero build step, instant reload.

Plugins can provide additional templates. Templates are HTML/CSS/JS scaffolds that get copied to the serve directory at startup:

browsercli start --template my-custom-template

Plugins can expose HTTP endpoints under the /x/ namespace. Handler scripts receive JSON on stdin and write JSON to stdout:

# handlers/refresh.sh
#!/bin/sh
INPUT=$(cat)
echo '{"ok": true, "refreshed_at": "'$(date -u +%Y-%m-%dT%H:%M:%SZ)'"}'

Call from client libraries:

// Node.js
const result = await ac.pluginRpc("/x/dashboard/refresh", { key: "value" });

# Python
result = ac.plugin_rpc("/x/dashboard/refresh", {"key": "value"})

Fire-and-forget scripts triggered by daemon events:

browsercli plugin init my-plugin   # Scaffold a new plugin
browsercli plugin list             # List installed plugins
browsercli start --template name   # Apply a plugin template at startup

All scripts receive environment variables: BROWSERCLI_TOKEN, BROWSERCLI_HTTP_PORT, BROWSERCLI_DIR, BROWSERCLI_BASE_URL, BROWSERCLI_STATE_DIR, BROWSERCLI_PLUGIN_NAME.

See PLUGINS.md for the full development guide, manifest schema, security model, and cross-platform notes. A complete example plugin is included.

1. browsercli start spawns a daemon process that:

   • Starts an HTTP static file server on a random port
   • Launches a Chromium browser via CDP (Chrome DevTools Protocol)
   • Opens a Unix socket (macOS/Linux) or TCP localhost (Windows) RPC server for CLI communication
   • Watches the served directory for file changes
   • Writes session state to ~/.browsercli/session.json (macOS/Linux) or %LOCALAPPDATA%\browsercli\session.json (Windows)

2. Subsequent CLI commands (goto, eval, dom, etc.) connect to the RPC endpoint and send JSON requests to the daemon.

3. The daemon translates RPC requests into CDP commands over WebSocket.

                             Unix Socket (macOS/Linux)
+-----------+           or TCP localhost (Windows)      +----------+
|  CLI cmd  | --------------> +--------------+ CDP/WS   | Chromium |
|           | <-------------- |    Daemon    | -------> |          |
+-----------+   JSON RPC      |              | <------- +----------+
                              | HTTP Server  |
                              | File Watch   |
                              +--------------+
                                     |
                                     v
                                Local Files

A zero-dependency Node.js client (written in TypeScript, ships with type definitions) is included in clients/node/. Install it with:

cd clients/node && npm install

import { BrowserCLI } from "@justinhuangcode/browsercli";

const ac = BrowserCLI.connect();   // reads ~/.browsercli/session.json
await ac.goto("/");
const title = await ac.domQuery("h1", "text");
await ac.screenshot("", "page.png");
await ac.stop();

// Plugin support
const plugins = await ac.pluginList();
const result = await ac.pluginRpc("/x/my-plugin/action", { key: "value" });

See clients/node/README.md for the full API reference.

A zero-dependency Python client is included in clients/python/. Install it with:

pip install -e clients/python

from browsercli import BrowserCLI

ac = BrowserCLI.connect()   # reads ~/.browsercli/session.json
ac.goto("/")
title = ac.dom_query("h1", mode="text")
ac.screenshot(out="page.png")
ac.stop()

# Plugin support
plugins = ac.plugin_list()
result = ac.plugin_rpc("/x/my-plugin/action", {"key": "value"})

See clients/python/README.md for the full API reference.

End-to-end examples are provided in examples/ for both Node.js and Python:

Run any example after starting the daemon:

browsercli start --dir /tmp/demo-site

# Node.js (build the client first)
cd clients/node && npm run build && cd ../..
node examples/01_write_reload_screenshot.mjs

# Python
python examples/01_write_reload_screenshot.py

browsercli stop

src/
├── main.rs              # CLI entry point and command dispatch
├── cli/mod.rs           # Command-line argument definitions (clap)
├── daemon/
│   ├── mod.rs           # Daemon module exports
│   └── server.rs        # Daemon process, RPC handler, session management
├── browser/
│   ├── mod.rs           # Browser module exports
│   ├── controller.rs    # CDP communication, browser lifecycle
│   ├── devtools.rs      # DevTools HTTP API client
│   └── find.rs          # Chromium/Chrome binary auto-detection
├── web/
│   ├── mod.rs           # Web module exports
│   ├── server.rs        # Static file handler with index resolution
│   └── welcome.rs       # Welcome page HTML template
├── rpc/
│   ├── mod.rs           # RPC module exports
│   ├── types.rs         # Request/response type definitions
│   └── client.rs        # RPC client (Unix socket / TCP)
├── plugins/
│   ├── mod.rs           # Plugin manifest types, validation, discovery
│   ├── registry.rs      # Central plugin registry with O(1) lookups
│   ├── executor.rs      # Script execution engine
│   ├── hooks.rs         # Lifecycle hook dispatch
│   └── templates.rs     # Template copy logic
└── watch/
    └── mod.rs           # File system watcher with debounce
clients/node/                # Node.js client library (TypeScript, zero dependencies)
clients/python/              # Python client library (zero dependencies)
examples/                    # End-to-end example scripts
examples/plugins/              # Example plugins (dashboard)
tests/
├── cli_integration.rs       # CLI integration tests
└── e2e_integration.rs       # Full lifecycle E2E test (requires Chromium)

browsercli is designed for single-user, local-only use on development machines. The following controls are in place:

• Multi-user / shared machines — Other local users with root or same-UID access can read the session token and issue RPC commands. If you run on a shared server, restrict access to ~/.browsercli/ via OS-level permissions or containers.
• Serving untrusted content — The HTTP server is intended for local files authored by you or your agent. Do not point --dir at untrusted directories.
• Production workloads — browsercli is a development/testing tool. It does not implement TLS, rate limiting, or audit logging.

By default, browsercli removes the navigator.webdriver flag and applies minor automation-fingerprint mitigations so that local pages behave as they would in a normal browser (e.g., some front-end frameworks alter behavior when they detect headless/automated Chrome).

Scope: Stealth mode is strictly for local development and testing where automation detection interferes with page behavior. It is not designed for bypassing security controls on external websites.

See TROUBLESHOOTING.md for common issues and solutions, including:

• Browser not found — install instructions per platform
• Port conflicts
• Headless mode on servers / CI
• Permission errors
• Template not found

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

See CHANGELOG.md for release history.

Inspired by steipete/canvas.

MIT