MCP tools

Jump serves these tools to your AI client over MCP. Each tool takes a node name from list_nodes.

Tool	What it does	Limits
list_nodes	Lists paired nodes.	—
node_info	Connects and reports hostname, OS, architecture, agent version.	30 s
run	Runs a command through the node’s shell (cmd.exe /C on Windows, /bin/sh -c elsewhere) and waits for it. Returns exit code, stdout, stderr.	Timeout 60 s by default, 30 min max. 256 KiB per stream.
read_file	Reads a file, paged by byte offset.	256 KiB per call by default, 4 MiB max.
write_file	Writes text to a file. Fails if the file exists unless mode is overwrite or append; create_dirs creates missing parent directories. Paths must be absolute. Does not set permissions.	Sent in 4 MiB pieces; if a later piece fails, the error says the file is incomplete.
tail_file	Returns the last N lines plus a cursor. Passing the cursor back returns only complete lines written since, so the AI can follow a log while a program runs. Detects truncation, and rotation when the new file is still smaller than the cursor.	100 lines by default, 5,000 max. 256 KiB per call.
start_process	Starts a long-running program (dev server, watcher, slow test run) through the node’s shell and returns its id at once. The agent captures its combined stdout and stderr.	32 processes per agent; the most recent 1 MiB of output per process.
process_output	Returns output since a cursor (or the last N lines) and whether the program is still running. Can wait for new output instead of being called repeatedly.	64 KiB per call by default, 256 KiB max. Waits up to 60 s.
stop_process	Stops a program and everything it started: asks them to exit (SIGTERM), then kills them after a grace period. Windows kills at once.	Grace 5 s by default, 60 s max.
list_processes	Lists started programs, running or finished, with their ids and state.	—
screenshot	Captures one of the node’s displays as the logged-in user sees it and returns a PNG image, with the display’s size so the AI can ask for a region at full detail. Needs the agent in the desktop session; see Screenshots.	Scaled to 1568 px on the long edge by default, 8192 max. 4 MiB per image (scaled down further to fit).
ui_tree	Reads a window’s accessibility tree: its buttons, fields, menus and lists, with their names, values, states, positions and the actions each supports, each with an id. Reads the window in front unless given a window handle (the reply lists the other windows) or an element to expand. On macOS the front app’s menu bar is listed too. On Linux Wayland desktops elements have no positions; see UI automation.	300 elements by default, 2,000 max. Names and values cut at 200 characters. 20 s.
ui_action	Acts on an element by id: click, focus, set_value, toggle, expand, collapse, select, scroll_into_view. Uses the app’s accessibility interface, so it works on windows that aren’t in front; click falls back to a real mouse click for elements with no action of their own.	Ids last for the node’s next 4 trees. Reports an action still running after 5 s (e.g. a dialog opened) instead of waiting.
click, scroll	Clicks (left, right, middle, double) or turns the mouse wheel at a point on a display, in display pixels. For what the tree doesn’t reach.	Scroll 1 to 50 notches.
type_text, press_keys	Types text into whatever has focus, or presses key combinations such as ctrl+s (cmd+s on macOS).	10,000 characters; 50 combinations.

Nodes added while an MCP session is running are picked up immediately. If a node’s agent is too old for a tool, the tool says which node to update.