I've been building Bruster - an open-source browser automation tool written in Rust, designed for AI agents and scripts.
The goal is simple: fast startup, lower memory usage, and a straightforward way to control a real browser.
Native webviews, built-in MCP
Bruster uses your platform’s native webview:
- macOS: WKWebView
- Linux: WebKitGTK
- Windows: WebView2
The controller runs without Node.js and exposes two interfaces: MCP for compatible AI clients and JSON lines over stdin/stdout for scripts.
You can connect it to Claude Code, Cursor, or another MCP-compatible client.
What can an agent do?
- Navigate websites and inspect page content.
- Click elements, fill forms, and submit them.
- Capture screenshots.
- Manage tabs and persistent sessions.
The demo is a scripted MCP session showing navigation, form interaction, and screenshots.
How much faster and lighter?
Here’s one published Linux benchmark comparing Bruster with agent-browser 0.40.0 using full Chrome:
| Metric, median | Bruster | agent-browser |
|---|---|---|
| Launch to first loaded page | 219 ms | 799 ms |
| Total idle memory, PSS | 256 MiB | 523 MiB |
That run used Ubuntu 24.04 and ten fresh sessions per configuration against local test pages. Memory includes browser helper processes and Bruster’s Xvfb display; Xvfb startup is excluded from timing.
These are workload-specific results across different browser engines. The full benchmark report includes additional tools, workloads, and methodology.
Try it with Claude Code
Install on macOS or Linux, then register the MCP server:
curl -fsSL https://raw.githubusercontent.com/aliilyas/bruster/main/install.sh | sh claude mcp add bruster -- "$HOME/.local/bin/bruster" --mcp --quiet Linux requires WebKitGTK 4.1, plus xvfb and xauth when running without a desktop display. Platform setup and other MCP clients.
Then try:
Use Bruster to open https://example.com, inspect the page, and take a screenshot.
Current limitations include synthetic input events and no cross-origin iframe access.
Bruster is MIT licensed, and feedback is welcome—especially on real-world agent workflows and MCP compatibility.
