hwatu
Fast, interruptible verification browser for AI coding agents: 35 ms checks, pixel diffs, live human hand-off
What is it?
What it is
Hwatu is a lightweight, warm‑daemon WebKitGTK browser written in Rust that provides fast, single‑call verification primitives for AI agents and allows seamless human hand‑off of live sessions.
Why it exists
It was created to eliminate the latency and resource waste of cold libraries like Playwright, avoid the focus‑stealing and overhead of full‑featured browsers, and give agents a deterministic, low‑cost way to prove visual and functional correctness with a single tool call.
Who should use it
AI agent developersLLM application engineersFrontend automation engineersDevelopers using Claude Code or CursorLinux users looking for high-performance browser automation
Who should avoid it
Windows and macOS users (Linux-only)Developers requiring Chromium-specific renderingUsers needing heavy-weight browser features (extensions, sync, etc.)
How it works
A quick walkthrough in plain English
How hwatu works
Step 1 of 3
You interact with it
Open hwatu, send a request, or connect it to your stack.
Features
Advantages
- Fast verification pass (~35 ms median) with a single tool call
- Warm daemon design keeps engine and GPU context hot, eliminating cold starts
- Headless by default prevents focus stealing while agents run
- Small binary footprint (static binary + system WebKitGTK)
- No need to download or ship Chromium
- Deterministic animation handling reduces flakiness
- Human hand-off preserves session state for debugging
- Built-in CAPTCHA detection and structured wait
- Cross‑agent compatibility via MCP, CLI, or socket
- Linux‑only simplifies dependencies and packaging
Disadvantages
- Linux‑only; no native support for Windows or macOS
- Relies on WebKitGTK, not Chromium, which may affect rendering parity
- AGPL‑3.0 license may be restrictive for some projects
- Limited feature set compared to full Playwright or CDP APIs
- Requires system WebKitGTK 6.0 to be installed
- Potential learning curve for new users unfamiliar with WebKit
Installation
native
curl -fsSL https://raw.githubusercontent.com/hongnoul/hwatu/main/scripts/install.sh | bash hwatu setup # Or on Arch: yay -S hwatu # Or from source: cargo build --release
FAQ
How does hwatu improve agent performance compared to Playwright?
Hwatu operates as a 'warm daemon' rather than a library. While Playwright requires engine startup for each task or complex manual connection management, hwatu keeps the engine, GPU context, and WebView prewarmed. This allows a full verification pass (load, eval, screenshot) to complete in ~35ms with a single tool call, whereas a warm Playwright server can take ~341ms and requires multiple API calls.
Can I interact with a headless browser session if the agent gets stuck?
Yes. Unlike most tools where headless mode is fixed at launch, hwatu treats headless as a window property. You can use the `hwatu focus <id>` command to materialize the live session into your tiling window manager, allowing you to handle CAPTCHAs or complex UI tasks before handing control back to the agent.
What are the system requirements and installation methods?
Hwatu is currently Linux-only and requires `webkitgtk-6.0`. You can install it via a shell script (`curl -fsSL... | bash`), via `yay -S hwatu` on Arch Linux, or by building from source using `cargo build --release`.
How do I connect my AI agent to hwatu?
Hwatu supports multiple connection methods: via the Model Context Protocol (MCP), a native Unix socket for Jcode, or a simple CLI. For manual MCP configuration, add the following to your config: `{"mcpServers": {"hwatu": {"command": "hwatu", "args": ["mcp"]}}}`.
What kind of verification metrics does hwatu provide?
Hwatu provides advanced verification primitives including pixel-diff scoring (with match percentage, regions, and heatmaps), animation metrics (duration, easing, and velocity), and structured JSON output for page state, console logs, and network requests.
Featured in Videos
YouTube tutorials and walkthroughs for hwatu
Alternatives
Similar projects ranked by category, topics, and text overlap.
