aihawk 0.6.0

An AI agent with a real browser. You say what you want in plain language, it goes and does it on the actual web.

_{FEATURED IN}
Business Insider · TechCrunch · Semafor · Wired · The Verge · Vanity Fair · 404 Media

Two ways to use it

The only question is where the model comes from.

1. You already use an assistant that can run tools

Your assistant brings the model. You add this browser to it, and nothing changes about how you work.

Claude Code:

claude mcp add --scope user stealth -- uvx invisible-playwright-mcp

Codex:

codex mcp add stealth -- uvx invisible-playwright-mcp

Gemini CLI:

gemini mcp add --scope user stealth uvx invisible-playwright-mcp

Then ask your assistant, in the window you already have open:

Go to news.ycombinator.com and give me the top five titles.

Claude Desktop, Cursor, VS Code, Windsurf, Zed and Cline take a config file instead, and the file is not the same shape for all of them. Each one is written out in the server's README.

2. You don't, or you want to watch it work

We bring the interface, you bring an OpenRouter key. Chat on the left, the live browser on the right.

uvx aihawk ui --openrouter-key sk-or-...

Then open http://127.0.0.1:8765 and type the same thing.

Same patched Firefox behind both. AIHawk reaches it through that MCP server, over MCP, exactly as your assistant would.

Before either one

Python 3.11 or newer, on Windows (x86_64) or Linux (x86_64, arm64). macOS is not supported: the last engine build for it was firefox-20.

Both commands above start with uvx, so you need uv:

curl -LsSf https://astral.sh/uv/install.sh | sh              # Linux
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"   # Windows

The browser is a separate download of about a quarter of a gigabyte, and it does not arrive when you install either side. It arrives on the first request that needs a page, so your first instruction sits there for a while and a slow connection can time out with an error that says nothing about a download. Get it over with first, where you can watch it:

uvx invisible-playwright fetch

What to ask it

Anything that needs a browser rather than an API, and a person's judgement about what is on the page.

Go to <paste the URL>. One way, Milan to Lisbon, economy, one checked bag, one adult. Check every date from the 12th to the 16th of next month, one at a time, and read the cheapest fare for each day. The date field is a calendar widget, so click the days rather than typing them. If a date has no availability, say so. Do not guess a number.

It drives the page the way a person would: the pointer moves, keys are pressed, and it refuses to set a form field from JavaScript even when that would be quicker, because a page can tell the difference.

Options

• --openrouter-key Your key, or the OPENROUTER_API_KEY variable.
• --model An OpenRouter model id, or AIHAWK_MODEL. Defaults to z-ai/glm-4.6.
• --proxy Optional. http://user:pass@proxy.example.com:8080 or socks5://proxy.example.com:1080. Host and port are both required. The timezone, locale and egress follow it.
• --binary An engine binary you already have. It must be the build the seal pins, or startup refuses: this skips the download, not the version check.
• --seed An integer. Same seed, same browser identity, every run.
• --profile-dir A directory to keep the profile in, so logins and cookies survive restarts.
• --headed Show the browser window. The interface shows you the page anyway.
• --host, --port 127.0.0.1 and 8765. Changing the host exposes an interface that has no authentication.

A .env beside the command

Rather than retyping the key and the binary path, put them in a .env in the directory you run from:

OPENROUTER_API_KEY=sk-or-...
STEALTHFOX_BINARY=/path/to/firefox

It is read at startup, and on the way in it never overrides something already set, so the order is --flag > the environment > .env > the default. Only the directory you are in is read - there is no search upwards, so running from a subfolder cannot silently pick up a different key. The startup line names the variables it applied and never prints their values.

Passing --openrouter-key puts the key in your shell history, and on Linux in the process list. OPENROUTER_API_KEY in the environment or in a .env avoids both.

Either way it does not reach the browser process: it is removed from the environment the engine starts with, by name and by value, so a copy under a second name goes too. tests/test_key_isolation.py fails if that stops being true.

The wiki

The reading room around the agent lives in the wiki: the AI browser-agent landscape and its comparisons, what to check when an agent gets blocked, and what happened to OpenAI Operator, among others. Worked examples, transcripts and their outputs live in articles/.

The rest of the family

• invisible-playwright-mcp The MCP server from option 1. Tools only, no interface.
• invisible_playwright The engine, as a Python library, for writing code instead of prompts. The API is Playwright's.
• invisible_core Seed to fingerprint to preferences, proxy and geolocation.

Contributing

Issues and pull requests welcome on whichever of those the problem lives in. If you are not sure, open it here. See CONTRIBUTING.

When something fails on a page, say which step, what the page did, what the tool returned and which exit country you were on. "It got blocked" is not something anyone can act on.

Using it responsibly

This automates a browser under your control. Read the terms of the sites you point it at, respect their rate limits, and do not submit anything a human has not read.

License

MIT. Everything distributed before 2 September 2026 was released under AGPL-3.0 and stays under it.