Configuration

~/.config/go-mcp-computer-use/config.json:

{
  "log_level": "info",
  "mouse_speed": 500,
  "click_delay_ms": 100,
  "verify_bounds": true,
  "action_timeout_ms": 30000,
  "uia_warmup": true,
  "training_enabled": true,
  "prior_adjustment": true,
  "watcher_auto_start": false,
  "watcher_interval_seconds": 5
}

Field	Default	Description
log_level	info	One of: debug, info, warn, error
mouse_speed	500	Mouse movement speed
click_delay_ms	100	Delay between mouse down/up (ms)
verify_bounds	true	Validate coordinates against screen bounds
action_timeout_ms	30000	Max time (ms) for blocking operations
uia_warmup	true	Warm up UIA at startup (async) to avoid cold-start delay. Set false if clients timeout during init.
training_enabled	true	Enable auto-save training snapshots on every UI action. Set false to stop all background data collection (also controllable at runtime via set_config).
prior_adjustment	true	Apply learned element frequency/position priors to ONNX detection scores. Set false for raw YOLO output only.
watcher_auto_start	false	Auto-start the background watcher on server boot. Watcher polls screen every N seconds and saves frames for training.
watcher_interval_seconds	5	How often the watcher captures and analyzes the screen (if running). Also used as default when starting via set_config.
tool_denylist	[]	Array of tool names to remove from the MCP server entirely (case-insensitive). Denied tools are invisible to the AI agent. Configurable at runtime via set_config.
retention_days	0	Auto-prune training samples older than N days. Deletes both database rows and image files. Background pruner runs every 6 hours. Set 0 to disable (default). Requires training_enabled: true. Configurable at runtime via set_config.

Transformer Model (Auto-Discovery)

The Go-native transformer engine requires no configuration. On server startup:

1. If model.gob exists in the data directory → loads it automatically
2. If no model exists → trains from the training_pairs table in SQLite
3. After each session → new training pairs are available for the next training cycle

No config fields needed — the transformer is self-managing.

Privacy Controls

See ../security.md for the full data collection and privacy controls reference.