-
-
Notifications
You must be signed in to change notification settings - Fork 4.6k
running aihawk with cline
Cline, the open-source coding agent that lives in VS Code, takes its MCP
servers as JSON entries under an mcpServers key, edited from inside the
extension: the MCP Servers icon in Cline's top toolbar, then the Configure
tab, then the Configure MCP Servers button, which opens the settings file for
editing. That path, and the field names an entry takes, are from Cline's own
MCP documentation. The block to paste is the same one every JSON-configured
client uses and it lives in the
server's README,
which this page links rather than copies, for the reason the
Claude Desktop page gives: a config
duplicated into a wiki is a config that rots in one of the two places.
The platform boundary, stated before you spend time: AIHawk's engine ships for Windows (x86_64) and Linux (x86_64, arm64), with no macOS build. VS Code and Cline run happily on a Mac; the server they start there would have no engine to run. This combination works on Windows and Linux today.
Paste the README's block under mcpServers in the file the Configure tab
opens, and Cline gains the browser as a set of tools. A server entry in that
file carries a command and args (here: uvx running
invisible-playwright-mcp), an optional env map, and two Cline-side
fields worth knowing from day one: disabled, which switches a server off
without deleting its entry, and autoApprove, a list of tool names allowed
to run without asking you each time.
The tools arrive in the same two families every other client sees: session
tools for tabs (session_new_page and friends) and page tools
(browser_navigate, browser_read_text, browser_snapshot,
browser_click, browser_type, browser_take_screenshot, among others).
Cline decides when a tool is relevant to your request and asks your approval
before running it; you describe outcomes and approve steps rather than
invoking anything by name. Cline's docs advise limiting autoApprove to
safe tools, and on a browser that acts on the real web the reading tools are
the sane candidates while the acting ones stay gated - the same
keep-the-consequential-click-human position
the forms page argues for everything
form-shaped.
Nothing else changes. There is no new account and no key on the AIHawk side: whatever model you already run Cline on does the thinking, the server brings only the browser, and the config block carries no secret. Nothing launches eagerly either; the browser exists from the first instruction that needs a page.
Cline's center of gravity is your codebase, so the browser earns its place the same way it does in Cursor: testing your own app through a realistic browser, in the same conversation as the agent that has your code open. "Open the local dev server, walk the signup flow, then look at the handler and explain the 500" is one request here. The worked version of that pattern, including what agent testing catches and where it is honestly flaky, is on the website-testing page. The second use is research that needs driving rather than fetching - pages built by JavaScript, walks through paginated docs - which the research page maps against cheaper tools.
For a plain lookup Cline's normal abilities already cover, skip the browser; a session plus model turns is the slow path, and it should be spent where acting on a page is the point.
The first prompt is the installation test, so make it small and checkable:
Open https://books.toscrape.com/ and tell me the title and price of the first book on the page.
One approved navigation, one read, one grounded answer, and you have verified the whole chain: config parsed, server started, engine present, page loaded. The site is a sandbox built for practice.
Then the Cline-native one, against something you own:
Start from http://localhost:3000. Try to register a new user with placeholder data, stop before the final submit, and list every validation message you saw. Then open the signup handler in this repo and tell me whether the messages match what the code enforces.
That second half is the reason to have a browser in a coding agent at all: the observation and the code review happen in one context.
The browser is headless by default; the agent's screenshots
(browser_take_screenshot) are your window. To watch it drive - worth doing
at least once against your own app - set STEALTHFOX_HEADLESS=0 in the
server entry's env map. The README documents the rest of the environment
variables; a persistent profile directory and a fixed identity seed are the
two most useful for repeated testing sessions.
-
The first page-touching prompt stalls. The engine is a separate download of about a quarter of a gigabyte, fetched on the first tool call that needs a page, not at install. Inside an editor that reads as the agent hanging on an approved step, and a slow connection can turn it into a timeout that never mentions a download. Prefetch it once, in a terminal where you can watch:
uvx invisible-playwright fetch
The engine is cached and shared with every other way into AIHawk.
-
The server never appears. Check the JSON you pasted (a trailing comma is the classic), and check that
uvxresolves for the process VS Code runs: the block launches the server withuvx, so uv must be installed where the editor finds it, which one working shell does not guarantee. Thedisabledfield is also worth a glance before deeper debugging. -
Every step asks permission. That is the design default, and while you are learning what the browser does, keep it. Loosen deliberately by adding read-only tools to
autoApproveonce the pattern is boring; Cline's own security note points the same way. -
It is a Mac. No engine build, per the boundary at the top; nothing to debug.
-
A public site loads oddly or pushes back. Separate the layers before touching config: browser problem or model problem for the local half, why agents get blocked for the site half.
How do I add AIHawk's browser to Cline? MCP Servers icon in Cline's
toolbar, Configure tab, Configure MCP Servers, then paste the block from the
server README under
mcpServers and save. No key, no signup; Cline's model does the thinking.
Does Cline use the browser automatically? It picks tools it judges
relevant, and by default asks approval per run. The autoApprove list makes
chosen tools run unasked; keep it to reading tools while the browser is new.
Do I need an API key for the browser? No. The server's block carries no secret and there is nothing to sign up for; your existing Cline model setup is untouched. The OpenRouter key belongs to AIHawk's own interface, a different way in, covered in using AIHawk without an API key.
Why is the first instruction so slow? Engine download: about a quarter
of a gigabyte on the first call that needs a page, silent from the editor's
side. uvx invisible-playwright fetch in a terminal moves that wait to a
moment you choose.
Can I watch it drive? Headless by default, screenshots as the window.
Set STEALTHFOX_HEADLESS=0 in the server entry's env to get a real
window, which is genuinely useful when it is testing your own app.
Does this work on a Mac? Not today: Cline runs there, the engine does not. Windows and Linux are the working platforms.
All retrieved 2026-09-03.
-
Cline docs: configuring MCP servers,
for the Configure-tab route, the
mcpServersJSON shape, thedisabledandautoApprovefields, and the limit-autoApprove security advice, with the MCP overview for the transport picture. - feder-cr/invisible-playwright-mcp, the server's README, for the config block, the tool list, the environment variables and the engine-download behavior.
- feder-cr/AIHawk, plus its README in this repository, for the platform boundary and the shared engine cache.
See also: running AIHawk with Claude Code, running AIHawk's browser from Claude Desktop, running AIHawk's browser from Cursor, and using an AI agent to test your own website.
From the AIHawk wiki. Fourth client, same block, same README: the config canon lives in one place on purpose, and this page is the tour around it, not a copy of it.
Alternatives and Comparisons
- OpenAI Operator alternatives
- Open-source Operator-style agents
- Is OpenAI Operator still available?
- OpenAI Operator vs Claude computer use
- browser-use alternatives
- Choosing an AI browser agent
- Open-source AI browser agents
- Open-source computer-use agents
- What is an AI web agent?
- AI browser agents vs traditional scraping
- Cloud browser infrastructure for AI agents, explained
- Browserbase alternatives
- Firecrawl vs an AI browser agent
- Skyvern alternatives
- Stagehand vs browser-use
- Project Mariner is gone: what replaced it
- Manus alternatives
- Gemini computer use vs Claude computer use
- AIHawk, reviewed honestly by its own wiki
- AI browser vs AI browser agent: which one do you want?
When the Agent Gets Blocked
- Why does my AI agent get blocked?
- The timing signal AI agents give off
- Agent retry loops trip rate limits, not fingerprints
- Claude computer use detected as a bot
- browser-use getting blocked: what you can and cannot change
Using the Agent
- Getting an AI agent to fill out forms
- Which model to use with AIHawk
- Browser problem or model problem?
- Running AIHawk's browser from Claude Code
- Using AIHawk without an API key
- Extracting data to a CSV with an AI agent
- Monitoring a page for changes with an AI agent
- Running AIHawk's browser from Claude Desktop
- Running AIHawk's browser from Cursor
- Using an AI agent to hunt for apartments
- Getting website data into Google Sheets with an AI agent
- Using an AI agent to download invoices from portals
- AI agents for web research
- Using an AI agent to test your own website
- Running AIHawk's browser from Cline
- Posting to social media with an AI agent
- Posting to Facebook with an AI agent
- Posting to Instagram with an AI agent
- Posting to X with an AI agent
- Automating LinkedIn posts: read this first
- Appointment bots: what they are and what an agent can legitimately do