Quirna MCP server

@quirna/mcp

Ask a human before your agent acts. Official Quirna server for the Model Context Protocol.

Add one block of config and your agent gains a tool it can call before doing something consequential — a refund, a production deploy, a destructive migration. A named human approves it on their phone. The agent proceeds only if they said yes, and the decision is signed and kept.

Nothing moves until someone says yes.

Your agent's framework may already have a way to pause for confirmation. This is different in three ways that matter once it is not just you at the terminal:

on their phone, behind Face ID — not to whoever happens to be watching the agent run.

• The decision leaves the loop. It goes to whoever your policy nominates,

asked live in your organization's settings. The agent cannot choose its own approvers, and neither can a prompt injection.

• Who approves is policy, not prompt. Quorum, risk tier and which group is

and exports to a file a third party can verify without trusting us.

• It leaves evidence. Every decision is signed at the moment it is made,

Install

Nothing to install: the config below runs it on demand with npx. To pin it, npm install -g @quirna/mcp and use quirna-mcp as the command.

Before you start

The mobile app is in private beta. Decisions are made on the phone and

nowhere else — the Console configures policies and never decides one — so

without the app an approval request is created and then waits until it

expires. Write to hello@quirna.com and we will get you in. Everything

else below works today.

Then, in the Console:

You get a key (ck_…), shown once.

1. A System key. Open Users → Systems and register one for this agent.

many of them are needed.

1. A Policy, which decides who approves a given kind of request and how

1. The app, on the phone of whoever will approve.

Configure your agent

Claude Code — claude mcp add quirna --env QUIRNAAPIKEY=ck_… -- npx -y @quirna/mcp

Cursor (~/.cursor/mcp.json), Claude Desktop (claudedesktopconfig.json) and most other clients take the same shape:

{
  "mcpServers": {
    "quirna": {
      "command": "npx",
      "args": ["-y", "@quirna/mcp"],
      "env": {
        "QUIRNA_API_KEY": "ck_your_system_key",
        "QUIRNA_REQUESTER_NAME": "Ops Agent",
        "QUIRNA_ENVIRONMENT": "production"
      }
    }
  }
}

Environment

On the key. It lives in your agent client's config file, which is often in a dotfile and sometimes in a repo. Treat it as a deployment credential: give each agent its own System rather than sharing one, scope it in the Console to the kinds it may request, and rotate it there if the file gets somewhere it should not be. A System key can only ask — it can never approve anything.

The tools

request_approval

Creates the request and waits for a human.

Shortened. The full README is on GitHub.