Exploring a screen by its pixels

You drive the person's real screen through the robomotion-image-mcp server, step by step, the way the robot will. On a Remote Desktop or Citrix session there is no element tree: the remote application arrives as a picture. The server reads that picture (OCR and box detection), you act on what it found, and every action is recorded as the Robomotion.ImageAutomation node that repeats it, with a template image cut around the target and verified unique on the whole screen with the package's own matcher - then the action itself is performed by the package's own code with that template. When you are done you get a sequence; creating-flow turns it into main.ts.

This is the pixel twin of exploring-browser, exploring-windows and exploring-java. Prefer those when they can see the application: a browser page (exploring-browser), a local Windows app (exploring-windows), a local Java app (exploring-java). Use this one when the application lives behind a remote session or draws its own controls.

Step 0 - load the tools

The mcp__image__* tools are deferred in Claude Code. Calling one cold sends malformed JSON and loses the server for the session. Before the first call:

ToolSearch query="select:mcp__image__image_status,mcp__image__image_list_windows,mcp__image__image_focus_window,mcp__image__image_snapshot,mcp__image__image_click,mcp__image__image_type,mcp__image__image_send_keys,mcp__image__image_find_text,mcp__image__image_screenshot,mcp__image__image_end"

Add the others as you need them (image_launch, image_wait, image_read_text, image_read_table, image_template, image_element_at, image_hover, image_drag, image_scroll, image_select_copy, image_write_templates, image_settings, ...). If the server shows as failed in /mcp, the binary is missing or not on PATH: say so; do not guess coordinates.

Before you start: the session the robot will run in

A template is pixels. It matches only a screen drawn the same way: