Let AI do
the clickwork
Hand your agent the jobs that take a mouse and a keyboard. It gets its own Mac, Windows or Linux desktop, looks at the screen, clicks, types and runs commands until the work is done.
Every machine is yours alone, accepts no inbound connections and holds nothing but what your agent puts there. When you're done it is wiped or destroyed.
Your agent
MCP- Install the new Mac build of Acme and go through onboarding like a new user. Tell me if anything feels off.
- Starting a Mac mini for this.
- create_instance os: "macos" ready · logged in to the desktop
- exec command: "open ~/Downloads/Acme-2.4.dmg" exit_code 0
- screenshot PNG · 1280 × 800
- drag path: [[470, 392], [810, 392]] Acme copied to Applications
- click x: 757, y: 750
- click x: 640, y: 422
- type text: "qa@clickwork.ai"
- click x: 640, y: 486
- screenshot a Documents access prompt
- click x: 710, y: 452
- screenshot PNG · 1280 × 800
- Acme installs and onboards cleanly on macOS. One thing feels off: it asks for your Documents folder before it says why it needs it. Screenshots attached; releasing the Mac now.
- destroy_instance Mac wiped · billing stopped
Live view
Works with Claude Code, Codex, Cursor, VS Code, Gemini CLI and any MCP client.
Your agent
One URL and an API key. Nothing to install. Create a key and these come with it filled in.
Run in a terminal. Add --scope user to use it in every project.
claude mcp add --transport http clickwork https://clickwork.ai/mcp \ --header "Authorization: Bearer YOUR_CLICKWORK_KEY"
Customize › Connectors › Add custom connector. It works in Claude Code on the web too: turn it on for the session. On Team and Enterprise plans an Owner adds it under Organization settings › Connectors, and everyone shares the key. If the dialog has no Request headers section, your account cannot send a Clickwork key yet.
Name Clickwork Remote MCP server URL https://clickwork.ai/mcp Authentication No sign-in Request headers authorization · Bearer YOUR_CLICKWORK_KEY · Required
Add to ~/.codex/config.toml.
[mcp_servers.clickwork]
url = "https://clickwork.ai/mcp"
http_headers = { "Authorization" = "Bearer YOUR_CLICKWORK_KEY" }
Add to ~/.cursor/mcp.json, or .cursor/mcp.json in a project.
{
"mcpServers": {
"clickwork": {
"url": "https://clickwork.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_CLICKWORK_KEY" }
}
}
}
Add to .vscode/mcp.json. VS Code asks for the key once and keeps it.
{
"inputs": [
{
"type": "promptString",
"id": "clickwork-key",
"description": "Clickwork API key",
"password": true
}
],
"servers": {
"clickwork": {
"type": "http",
"url": "https://clickwork.ai/mcp",
"headers": { "Authorization": "Bearer ${input:clickwork-key}" }
}
}
}
Run in a terminal. Add -s user to use it in every project.
gemini mcp add --transport http clickwork https://clickwork.ai/mcp \ --header "Authorization: Bearer YOUR_CLICKWORK_KEY"
Any MCP client that speaks Streamable HTTP and sends a bearer token.
{
"mcpServers": {
"clickwork": {
"type": "http",
"url": "https://clickwork.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_CLICKWORK_KEY" }
}
}
}
MCP server
Files move by URL, not through your agent's context. The same actions are a REST API for scripts and CI.
Use cases
The clickwork a real desktop takes off your hands.
Test on Windows from a Mac
Your agent builds on macOS or Linux and the bug report says Windows. It starts a clean Windows Server, installs the build, clicks through the installer, UAC prompts included, and comes back with screenshots.
Check the actual screen
The tests pass and the window is still blank. Your agent opens the app on a real desktop and looks at it after every change, the way a person would.
Build on a real Mac
A dedicated Mac mini on Apple silicon, not a VM. Install Xcode, build, and run the simulator and the test suite.
Drive apps with no API
Legacy desktop software, admin consoles, installers: if a person can do it with a mouse and keyboard, your agent can do it from screenshots.
Try risky changes
Snapshot the disk before an upgrade or a migration and restore it in minutes if it goes wrong. On Windows and Linux.
Run code you don't trust
A new dependency, a script from an issue, an unknown installer: run it on a machine with no credentials and no route to your network, then destroy it.
One owner
Every machine is created for you and deleted after you.
-
01
Connect
Add the MCP server to your agent with one URL and an API key.
-
02
Start
Your agent calls create_instance. Linux is ready in about 90 seconds, Windows in about 3 minutes and a Mac in about 2½, logged in to the desktop.
-
03
Drive
Screenshots, clicks, keys and a shell until the job is done. Watch live in your browser and take control when you need to.
-
04
Destroy
Your agent calls destroy_instance, or the TTL runs out. Billing stops; Windows and Linux are deleted and Macs are wiped.
Machines
Prices in USD. Egress included.
macOS
macos-defaultWindows
windows-defaultLinux
linux-defaultTools
One MCP server for all three operating systems.
create_instance · list_instances
Linux, Windows Server or macOS, with a TTL so a forgotten session cannot run up a bill.
screenshot
The screen as a PNG, 1280 × 800. Coordinates are its pixels.
click · move · scroll · drag
Pointer input in screen pixels, hover states included.
type · key
Text and key combinations, including installer UAC prompts on Windows.
exec
A shell command as root with its exit code and output. Long jobs run in the background.
upload · download
Files up to 5 GB by URL, never through your agent's context.
snapshot · restore
Save the disk before a risky upgrade and roll back in minutes.
extend_ttl · usage
Time left and cost so far, any time.
destroy_instance
Billing stops. Windows and Linux are deleted; Macs are wiped.
Secure and isolated
Your agent clicks around on a machine nobody else uses, with nothing of yours on it but what it brings.
No inbound ports
The agent inside each machine dials out to the gateway. Nothing on a machine listens on the internet.
No shared network
Machines never share a private network and cannot reach the platform. They reach each other only as strangers on the internet.
No credentials
A machine holds no cloud keys. Its only secret is a 24-hour certificate for its own tunnel.
Questions
Which agents work with Clickwork?
Any MCP client that connects to remote servers over Streamable HTTP with a bearer token: Claude Code, Codex, Cursor, VS Code, Gemini CLI and others. Anything else can use the REST API with the same key.
Do I install anything?
No. Clickwork is a remote MCP server and the machines run in Clickwork's cloud. Your agent needs the URL and an API key.
How does my agent see the screen?
The screenshot tool returns the screen as a PNG, 1280 × 800 by default, and click, move, scroll and drag take coordinates in that image's pixels. For anything faster in a terminal, exec runs a shell command and returns its output.
How long does a machine take to start?
About 90 seconds for Linux, 3 minutes for Windows and 2½ minutes for a Mac, which is set up for you when you start it. A machine is ready when it is logged in to the desktop.
What if my agent forgets to clean up?
Every machine has a TTL: 2 hours unless you choose between 10 minutes and 7 days (a Mac's is never under 24 hours). When it runs out the machine is destroyed and billing stops. Your agent can check the time left and cost with usage and add time with extend_ttl.
Can I watch what my agent is doing?
Yes. Every machine has a live view in the dashboard. It is read-only until you take control, so you can step in at a sign-in or a two-factor prompt and hand back.
Can I SSH in or open a port?
No. Machines accept no inbound connections. Your agent reaches them through the tools, and you reach them through the live view.
How am I billed?
From prepaid credit in USD, bought by card through Stripe. Machines are charged per started hour; a Mac for at least 24 hours, charged up front, as Apple's licence requires. Egress is included.
Hand off the clickwork
Create an account, add credit, paste one line into your agent.