MCP Tools Reference

automata-agent exposes all ui-automata capabilities to Claude as MCP tools. Each tool covers a category of operations; within a tool, an action parameter selects the specific operation. This keeps the tool list short and the schema self-documenting.

Workflow Tools

workflow

Manage the workflow lifecycle.

Action	Description
list	List available workflow files
status	Get the status of the currently running workflow
cancel	Cancel the running workflow
list_runs	List recent run logs with elapsed time
lint	Validate a workflow YAML string without executing it

start_workflow

Run a workflow file to completion and return its outputs. Accepts a file path or an inline YAML string and a params map. Streams phase progress notifications back to the client while running.

This is the primary way an agent executes a predefined automation — hand it a workflow from the library and let the executor handle all the UI interaction.

run_actions

Execute a list of UI automation steps directly, without a workflow file. Each step has an intent, an action, and an expect condition — the same structure as a workflow step. The executor runs them against a live window and returns the results.

Useful for one-off interactions or for an agent composing its own steps on the fly rather than calling a pre-built workflow.

UI Inspection

desktop

Inspect the Windows UI element tree.

Action	Description
list_windows	List all top-level windows with HWND, title, process, and bounds
element_tree	Dump the full UIA element tree of a window; accepts an optional selector to filter to matched subtrees
find_elements	Run a live selector query against a window and return all matches with role, name, and bounds

The primary tool for an agent exploring an unfamiliar application before writing selectors or deciding how to interact with it. Works alongside vision — desktop gives the UIA structure; vision gives the visual layout.

vision

OCR and visual layout capture.

Action	Description
window_layout	Capture visible text and UI regions for a specific window
screen_layout	Full-screen capture
window_probe	Hover-probe mode — discovers interactable elements not exposed via UIA by simulating mouse movement

Use vision when UIA gives incomplete information — custom-rendered controls or applications that draw their own widgets. Coordinates returned by vision are in screen space and can be passed directly to input mouse_click.

Application Management

app

Launch and manage applications and windows.

Action	Description
list_installed	List installed applications
launch	Launch an application; wait strategy: new_pid, new_window, or match_any
list_taskbar	List windows pinned or open in the taskbar
focus	Bring a window to the foreground by HWND
show_desktop	Minimize all windows and show the desktop
list_task_view_windows	List windows visible in Task View
activate_task_view_window	Switch to a window via Task View

window

Manipulate a specific window by HWND.

Action	Description
minimize	Minimize the window
maximize	Maximize the window
restore	Restore to normal size
close	Send WM_CLOSE
set_bounds	Reposition and resize the window
screenshot	Capture a screenshot of the window

Get the HWND from desktop list_windows.

Input

input

Raw mouse and keyboard input. Operates at the OS level — works on any window regardless of UIA support.

Action	Description
mouse_move	Move the cursor to screen coordinates
mouse_click	Click at coordinates; button: left, right, middle, double, triple
mouse_drag	Click and drag from one point to another
mouse_scroll	Scroll at coordinates
key_press	Send a key or chord with modifier syntax ({ctrl}v, {alt}{F4})
type_text	Type a string as keystrokes
get_cursor_pos	Return the current cursor position

clipboard

Read or write the Windows clipboard.

Action	Description
read	Return the current clipboard text
write	Set the clipboard to a string

Useful for extracting text that is easier to copy than to read via UIA, or for injecting data into an application via paste.

Browser

browser

Control Microsoft Edge via the Chrome DevTools Protocol.

Action	Description
ensure	Start Edge with a CDP debug port (must call first)
tabs	List open tabs with id, title, and URL
navigate	Navigate a tab to a URL
eval	Evaluate JavaScript in a tab and return the result
dom	Read the DOM tree of a tab
screenshot	Capture a screenshot of a tab
open	Open a new tab
activate	Switch to a tab by id
close	Close a tab by id

For web-based workflows where UIA is insufficient — Edge's UIA tree does not expose full page content. CDP gives direct access to the page DOM and JavaScript runtime.

Filesystem

file

Full filesystem access.

Action	Description
list	List files in a directory
read	Read a file's full contents
read_lines	Read a specific line range from a large file
write	Write (create or overwrite) a file
append	Append to a file
copy	Copy a file
move	Move or rename a file
delete	Delete a file
mkdir	Create a directory
rmdir	Remove a directory
stat	Get file metadata (size, modified time, etc.)
glob	Search with a glob pattern
checksum	Compute a file checksum

Handles binary files (base64-encoded) and large text files (via read_lines).

System

system

Shell execution, process management, and system information.

Action	Description
exec	Run a program directly (no shell); pass program and arguments as an array
ping	Check connectivity to the agent
whoami	Return the current Windows username
ipconfig	Return network adapter configuration
route_table	Return the IP routing table
get_path	Return the current PATH environment variable
list_processes	List running processes
kill_process	Terminate a process by PID

cmd.exe and powershell.exe can be passed as the program to exec for shell-style invocation.

Library

resources

Browse the embedded workflow library.

Action	Description
list	List available workflows with their descriptions and parameters
read	Read a specific workflow file by path

An agent can use this to discover what pre-built workflows are available before deciding how to approach a task.