Lab 5: Through Canine's MCP tools

list_agent_computers only reads the database. The others look up the VM's launcher pod (~1.2 s) and open a new kubectl port-forward (~1.6 s) before the ~0.5 s action. Keeping the forward open between calls is the obvious fix (see open work).

The image content is what the model looks at. The text tells it the image's size in pixels, which is the coordinate space for its next action.

No. Both get "not found or you don't have access to it", so the tool doesn't reveal which IDs exist.

== 1. The four tools Canine's /mcp endpoint offers for agent computers, and what they take
$ tools schemas
list_agent_computers: List the agent computers (cloud desktops) you can use. Their IDs are what the computer_* tools take.
   inputs:    (required: none)
computer_screenshot: Take a screenshot of an agent computer's desktop. Coordinates in the image are what computer_action takes.
   inputs: agent_computer_id   (required: agent_computer_id)
computer_action: Use an agent computer's mouse and keyboard, following Anthropic's computer-use tool: e.g. left_click at a c...
   inputs: agent_computer_id, action, coordinate, start_coordinate, text, scroll_direction, scroll_amount, duration, region   (required: agent_computer_id, action)
computer_accessibility: Read and use apps on an agent computer by their structure instead of pixels, through the accessibility tree...
   inputs: agent_computer_id, operation, app, role, name, path, text, max_depth   (required: agent_computer_id, operation)
   computer_action's inputs are the fields of Anthropic's computer-use tool, so a model's action passes straight through.

== 2. Call them as the computer's owner, and time each call
$ tools calls
list_agent_computers()  # which computers can I use?
   [{"id":11,"name":"omarchy","status":"running","cluster":"devserver"},{"id":12,"name":"omarchy","status":"running","cluster":"cluster"},{"id":14,"name":"omarchy-auto","status":"running","cluster":"devserver"}]   (0.04s)
computer_screenshot(agent_computer_id: 14)  # what's on screen?
   [image image/png, 1148000 base64 chars] + {"width":1280,"height":720}   (4.96s)
computer_action(agent_computer_id: 14, action: "cursor_position")  # where's the mouse?
   {"coordinate":[640,371]}   (3.69s)
computer_action(agent_computer_id: 14, action: "mouse_move", coordinate: [400, 300])  # move it
   {}   (3.67s)
computer_action(agent_computer_id: 14, action: "zoom", region: [0, 0, 320, 25])  # look closely at the top bar
   [image image/png, 4252 base64 chars]   (3.84s)
computer_accessibility(agent_computer_id: 14, operation: "tree", max_depth: 0)  # which apps have a tree?
   {"apps":[{"path":"0","role":"application","name":"quickshell","states":["disabled"]},{"path":"1","role":"application","name":"xdg-desktop-portal-gtk","states":["disabled"]},{"path":"2","role":"application","name":"qui...   (3.7s)
computer_action(agent_computer_id: 14, action: "left_click", coordinate: [99999, 0])  # a mistake
   ERROR coordinate [99999, 0] is off the 1280x720 screen   (3.27s)
computer_screenshot(agent_computer_id: -1)  # someone else's (or no) computer
   ERROR Agent computer not found or you don't have access to it   (0.01s)
   Every call looks up the VM's pod (~1s) and opens a fresh kubectl port-forward (~1.5s) before the action runs.
   Screenshots come back as MCP image content, so Claude sees the picture, plus the size as text.