Connect an AI assistant (MCP)

Quiet Guard exposes a read-only MCP (Model Context Protocol) server. An assistant you connect to it can answer "which client is on fire this morning", "how long has this exception been coming back" or "which dependencies of this project have a security advisory", without you opening one tab per client.

It cannot change anything: not resolve an issue, not assign one, not touch a setting.

Before you start

Three conditions, in this order.

1. Your team is on a plan that opens the MCP server: Solo, Indie or Studio.

The Free plan does not open it.

1. A team owner has opened your seat, on the Members page. The switch is

closed by default, owners included: connecting a third-party assistant to the team's data is a decision, not a consequence of a role.

1. Your account is approved. During the approval beta, an account still waiting

reaches nothing, token or no token.

If one is missing, the assistant will say so in words rather than hand you an empty list.

The server address

https://quietguard.dev/mcp

Nothing after it, and no key to paste: authentication is OAuth and your client handles it. On the first connection, a Quiet Guard page asks for your consent; it names the application's real redirect domain, the name it declares, and the list of projects it will be able to read.

claude.ai and Claude Desktop

Settings, then Connectors, then Add custom connector. Paste the address above. Claude opens the consent page in your browser.

ChatGPT

Settings, Connectors, Add. Same address, same consent page.

Claude Code

claude mcp add --transport http quiet-guard https://quietguard.dev/mcp

Then /mcp in a session to start authentication. If the server does not show up, restart the application completely: the server list is read at startup.

Cursor

Settings, MCP, Add new global MCP server, then in the JSON:

{
  "mcpServers": {
    "quiet-guard": {
      "url": "https://quietguard.dev/mcp"
    }
  }
}

VS Code

Command palette, MCP: Add Server, HTTP, then the address.

What the assistant can ask for

Seven tools, all read-only.

Every tool takes a project key written team/project, the one list-projects returns. A project slug is only unique within its team, which is why the key has two halves.

One resource, quietguard://methodology, explains how the product counts: an issue is not an occurrence, a retention window explains a curve that stops, and a missing percentage is not a zero.

What the assistant cannot do

• Write anything. No tool modifies.
• Read your application logs. They are not in the catalogue.
• Read an occurrence's context. That is where request bodies live, and they

do not leave.

• Read encrypted content without your say-so. If your team turned on

encrypted storage, the assistant gets the metadata and the fact that the content is sealed, until you unlock it (see below).

Reading encrypted content

A team owner first allows the principle, on the Encryption page, AI assistants section. Then each member can choose Unlock for my assistants, type the passphrase and pick a duration: 15 minutes, 1 hour or 4 hours. For that time, the assistant reads the message and trace of issues; the occurrence context never leaves.

• The passphrase is typed on that page, never in the conversation. When

the content is sealed, the assistant is given the page's address.

• The unlock is yours alone, for every application you connected. Your

colleagues read nothing more.

• Lock now, withdrawing an application, your MCP seat closing, a

passphrase change or the owner withdrawing the permission stop it at once. Locking the dashboard or signing out does not.

• For the chosen duration, the server holds the team key, as with the

dashboard's session unlock.

Withdrawing access

Panel, account menu (your avatar, top right), Connected applications page. Revocation takes effect on the very next call.

Closing your seat's switch on the Members page has the same immediate effect, without touching anybody else. And losing the plan, or falling into dunning, closes the server for the whole team without a single token needing to be revoked: rights are never frozen into the token, they are read again on every call.