AIHawk: an open-source browser agent that speaks MCP
The MIT-licensed project drives a real browser from plain English, either as an MCP server for coding assistants or through a local web interface.
Symbolic illustration: a small robotic arm works a laptop trackpad while abstract color blocks scroll across two monitors.
AIHawk is an open-source agent that turns a plain-English instruction into real clicks and keystrokes in a live browser, running either as an MCP server for Claude Code and Gemini CLI or as a local web app.
At a glance
- GitHub at time of reading: 30.3k stars, 4.6k forks, 198 watchers, 225 commits on the main branch
- Licensed MIT, but everything distributed before September 2, 2026 stays under AGPL-3.0
- One command wires it into Claude Code as an MCP server using uvx aihawk
- The bundled web UI serves on 127.0.0.1:8765 and expects an OpenRouter API key
- Python is the primary language; browser control runs on Playwright via the invisible_playwright engine
AIHawk asks for a task in ordinary English and then carries it out inside a real browser, moving the cursor, clicking and typing rather than calling an API. At the time of reading, the repository showed 30.3k stars and 4.6k forks.
Plain English in, real clicks out
What separates this from a scraper is where it operates. The agent reads the rendered page and acts on it, instead of chasing brittle HTML selectors. Underneath sits Playwright, wrapped in the project's own invisible_playwright engine, with a companion package called invisible_core handling fingerprint, proxy and geolocation settings.
The main branch carries 225 commits and the repository lists 198 watchers, with Python as the primary language. Browser profiles and reproducible seeding are offered as configuration options.
Two front doors: MCP server or local web app
The first route is Model Context Protocol. A single line registers the tool with Claude Code (claude mcp add --scope user stealth -- uvx aihawk), after which the assistant can reach for the browser like any other tool. The project points Gemini CLI and other MCP-capable assistants at the same mechanism.
The second route skips the assistant entirely. A bundled UI listens on 127.0.0.1:8765 and needs an OpenRouter key, supplied on the command line or through OPENROUTER_API_KEY, with the model selected by flag or by the AIHAWK_MODEL variable. The Playwright binary is fetched automatically unless you point the tool at your own.
A license with a cutoff date
The headline license is MIT, with one caveat that matters to anyone shipping a product: code distributed before September 2, 2026 remains AGPL-3.0. Fork an older tree and the copyleft obligations travel with it; only the post-cutoff code is permissive. That single sentence is the most consequential line on the page for commercial adopters.
What the page does not answer
The project warns users not to submit anything a human has not read, and to stay within the terms of service and rate limits of the sites being visited. It publishes no benchmark or success rate for how well the agent handles real websites.
Three things could not be confirmed from the page and are not asserted here: no version number or release tag was visible, the contributor count was not shown, and nothing on the page documents an earlier purpose or a rename of the repository.
FAQ
What does AIHawk actually do?
It takes an instruction written in plain English and performs it in a live browser, navigating, clicking, typing and reading what is on the page.
Do I need an API key to run AIHawk?
For the standalone web UI, yes: it expects an OpenRouter key via a flag or the OPENROUTER_API_KEY variable. Run as an MCP server, your assistant supplies the model.
Is AIHawk MIT or AGPL licensed?
Both, depending on vintage. Current code is MIT, but anything distributed before September 2, 2026 remains under AGPL-3.0.
