Locator system prompt (structured tools — Claude or local model)
Pixel system prompt (screenshot computer-use — Claude)
Vedon locators
structured UIA — no screenshots
idle
0.0s
elapsed
0.0s
thinking
0.0s
tool exec
$0
cost
0
tokens (in/out)
0
tool calls
0
turns
Start a session, then Run
💬 Activity — thinking · tools · output idle
Run to see the agent's thinking, tool calls, and output…
Screenshot computer-use
vision + coordinate clicks (pixels only)
idle
0.0s
elapsed
0.0s
thinking
0.0s
tool exec
$0
cost
0
tokens (in/out)
0
tool calls
0
turns
Start a session, then Run
💬 Activity — thinking · tools · output idle
Run to see the agent's thinking, tool calls, and output…
Run showdown auto-creates two anonymous throw-away sessions (each a distinct
local user, so they run side-by-side even on a single-session host) and starts both scenarios. Each panel picks
its own model (the locator side can run a LOCAL model — llama/phi via Ollama — for free) against the same task;
the tool surface differs (structured locators vs screenshot pixels). Tool exec time is harness-measured
(dispatch→result); thinking is the model's wall-clock reasoning. The activity log under each panel streams
the agent's thinking, tool calls, and output live.