# MCP security

> What protects you when an AI assistant uses SmartChat. Permissions, server checks, untrusted text and how to cut access.

Last updated: 2026-10-30. Source: https://docs.smartchat.marketing/en/mcp/security/

An AI assistant can make mistakes or be misled by text it reads. SmartChat limits the damage in several layers.

| Layer | What it does |
|---|---|
| Permissions | The assistant only gets the tools you agreed to. The consent page starts with **Read**. |
| Separate ticks | Sending, publishing, submitting templates, reading secrets and running app actions are ticked on their own. |
| Server checks | Opt-in, unsubscribed contacts, the 24-hour window and approved templates are checked on the server for every call. See [Sending safely](https://docs.smartchat.marketing/en/security/). |
| Untrusted text | Text from contacts is marked with `_untrusted`. See [Untrusted content](https://docs.smartchat.marketing/en/concepts/untrusted-content/). |
| No dangerous tools | No deleting, no plan or payment, no account or team management, no key management. |
| Repeat protection | Calls that send are protected against double sends. |
| Audit | Calls that change or send are written to an audit log. |
| Protection limit | 120 requests per minute for each API key or MCP connection. REST and MCP calls with the same key or connection count together. |

## Cut access

Revoke the connection or the key in the SmartChat app. It stops working at once.