---
title: "Think-Act Security And Privacy | Desktop AI Guardrails"
description: "Learn how Think-Act controls desktop AI risk with tool permissions, approval gates, disabled failed connectors, logs, local endpoints, and security policy."
---

# Think-Act Security And Privacy

Think-Act is designed for desktop work, so security cannot be an afterthought. A desktop agent may read files, use tools, process documents, or call local MCP servers. Those actions need clear controls.

## Security Principles

The product follows practical rules.

- The user decides which tools are enabled.
- Risky actions can require confirmation.
- Failed tools are not exposed to the agent.
- Skills cannot replace missing connectors.
- Logs should help users understand what happened.
- Sensitive workflows should use approved endpoints.

## Tool Permissions

Connectors expand what the agent can do.

- Enable only trusted MCP servers.
- Disable tools that are not needed.
- Keep write and delete actions behind confirmation.
- Block commands or tools that violate policy.
- Remove local servers that fail or behave unexpectedly.

## Local MCP Risk

Local MCP packages can run code on the device.

- They may need Node.js, npx, Python, or other runtimes.
- They may access files, networks, credentials, or app data exposed to them.
- Third-party packages are not controlled by PixLab.
- Users should verify package source, permissions, and environment variables.
- Trust confirmation is required before installation.

## Data Handling

Security depends on where data is processed.

- Use local or private endpoints for sensitive documents where configured.
- Avoid sending secrets, tokens, private customer data, or regulated data to unapproved tools.
- Review extracted OCR content before sharing.
- Keep recordings focused and avoid private screens.
- Follow organization policy for retention and storage.

## Approval And Recovery

Mistakes need safe exits.

- Require approval for risky commands.
- Inspect tool observations after runs.
- Use rewind and checkpoint controls where available.
- Review changes before trusting the final answer.
- Keep high-impact actions manual until proven safe.

## Accessibility And Safety

Security controls should remain usable.

- Warnings should be readable in light and dark themes.
- Critical prompts should be clear and localized.
- Keyboard workflows should stay accessible.
- High-contrast support should not hide security state.
- Error labels should explain what failed.

## Enterprise Direction

Future enterprise features should strengthen control.

- Managed installers and update policy.
- Admin-approved connector registry.
- Cloud MCP connector gating.
- A2A remote agents after connector security is proven.
- Windows least-privilege worker account planning for risky tools.

## Get Started

Use security defaults before adding tools.

[Download Think-Act now](./downloads.md) and begin with writing, OCR, and screen recording before enabling local MCP packages.

## Related Pages

- [Home](./index.md)
- [Integrations](./integration.md)
- [Desktop Agent](./feature-desktop-agent.md)
- [Accessibility](./accessibility.md)
- [FAQ](./faq.md)
- [Downloads](./downloads.md)

## FAQ

### Are MCP tools safe by default?

No. Local MCP servers can run code. Install only trusted packages and keep risky permissions behind confirmation.

### Can Think-Act use private endpoints?

Yes, where configured. Sensitive workflows should use approved local or private endpoints.

### What happens when a connector fails?

It should be disabled and excluded from available tool calls until the user fixes and re-enables it.