MRCO AI logoMRCO AIWINDOWS INTELLIGENCE
Security, privacy & control

Powerful tools should remain understandable.

MRCO AI can interact with real apps, files, browser sessions and commands. Understand what stays local, what can be sent to providers, where configuration is stored and how to reduce risk before you automate.

MRCO AI / SECURITY, PRIVACY & CONTROLSYSTEM READY
PERMISSIONS MATTERREVIEW → EXECUTE
Practical transparency

Know the boundaries before you automate

Local inference, local storage and remote actions are different things. A local model can still use a web tool, and cloud modes send request context to configured providers.

01◈

Local data store

Chat history, saved memory, knowledge chunks, embeddings, skill files and settings live under the app data directory on the Windows user profile.

Review guidance ↗
02⇄

Provider data transfer

Online mode sends prompts and assembled context to the configured provider. Hybrid may send parts of a request to multiple model providers.

Review guidance ↗
03⌘

Real-world side effects

Commands, browser clicks, key presses, scripts and file operations can change real system or website state.

Review guidance ↗
04⚿

API key handling

The current documented implementation stores provider settings and API keys in a local JSON file without encryption in that code path. Protect the directory.

Review guidance ↗
05◉

Screen and voice privacy

Vision captures a screen or window for analysis; speech features use the selected recording device and voice engine. Avoid sensitive content.

Review guidance ↗
06✓

Human oversight

Start with read-only requests, constrain paths, review elevated commands, inspect custom scripts and confirm consequential actions.

Review guidance ↗
Before a task runs

Guardrails are a workflow choice.

MRCO AI includes per-tool approval behavior for some consequential actions—for example, file deletion and certain dangerous UI interactions. Because not every system change necessarily follows an identical confirmation path, users should treat commands and scripts as real execution.

  • ✓Start with read-only inspection and previews.
  • ✓Verify the path, target app and destination before file changes.
  • ✓Understand a command before requesting elevated privileges.
  • ✓Review custom scripts before installing or executing them.
  • ✓Use least-privilege API keys for market-data experiments.
SAFER EXECUTION CHECKLIST

[1] Inspect target and current state
[2] Present plan and expected changes
[3] Wait for approval if risk warrants it
[4] Execute the scoped action
[5] Verify result and report evidence

REMEMBER: A confirmation prompt is not a substitute for understanding the action.
Data map

What lives on this PC?

The guide identifies the application data root as %LOCALAPPDATA%\AiDesktop. The exact contents depend on installed features and how you use them.

AreaTypical contentsCare point
settings.jsonMode, language, selected models and behavioral options.Keep configuration in trusted local storage.
online.jsonProvider endpoints, models and configured API keys.Keys are not encrypted by the documented code path.
memory.dbHistory, memory, actions, knowledge chunks and embeddings.Back up deliberately; protect the user profile.
models / runtimeLocal models and runtime assets.May consume significant disk space.
browser-profileSeparate automation browser profile.Web sessions/data can remain stored locally.